Ocean engineering data management system and method based on machine learning

Abstract

The invention discloses an ocean engineering data management system and method based on machine learning, and relates to the technical field of ocean engineering equipment fault diagnosis. The system comprises a multi-dimensional sensor array deployment module, a noise feature topology library management module, a real-time noise separation and feature extraction module, a fault rule chain deduction module, a multi-dimensional fault judgment module and an intelligent operation and maintenance feedback module. A sensor is arranged at a stress concentration part of equipment to collect vibration data, a fault feature vector is extracted after phase-space reconstruction and noise separation, then a rule chain is used for deducing a fault type and a fault position, and finally a structured maintenance report is generated and pushed. According to the system, the problems of multi-source noise interference, insufficient positioning precision and the like in the marine environment are solved, the residual noise energy proportion after noise separation is smaller than 5%, the fault position error is smaller than or equal to 0.1 mm, the comprehensive misjudgment rate is reduced, and the operation and maintenance efficiency and reliability of marine engineering equipment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine engineering equipment fault diagnosis, and particularly relates to a marine engineering data management system and method based on machine learning. BACKGROUND

[0002] Marine engineering equipment is long-term served in extreme environments such as high salt mist, strong corrosion and complex sea conditions, and its fault diagnosis and operation and maintenance face multiple technical bottlenecks: Multi-source noise interference: traditional vibration monitoring methods are difficult to accurately extract equipment fault features from background noise such as sea current impact and tidal fluctuation, resulting in a misjudgment rate of more than 30%.

[0003] Environmental parameter coupling effect: the running state of the equipment is dynamically disturbed by parameters such as sea current velocity and tidal phase, for example, when the sea current velocity is greater than 2 m / s, the bearing vibration amplitude threshold needs to be floated by 20% to effectively identify faults, but the existing system lacks a dynamic threshold correction mechanism.

[0004] Insufficient spatial positioning accuracy: the spatial error between the fault location and the vibration source often exceeds 0.5 mm, which cannot meet the precise maintenance needs of marine engineering equipment, for example, if the positioning error of the loose blade tenon is greater than 0.1 mm, it may cause chain structure damage.

[0005] Lagging nature of operation and maintenance decision: the traditional process relies on manual analysis reports, and the average time from fault discovery to maintenance execution is more than 24 hours, which is difficult to cope with the high economic loss caused by the shutdown of deep sea equipment. SUMMARY

[0006] The purpose of the present application is to provide a marine engineering data management system and method based on machine learning to solve the problems in the prior art.

[0007] To achieve the above purpose, the present application provides the following technical solutions: The marine engineering data management system based on machine learning comprises a multi-dimensional sensor array deployment module, a noise feature topology library management module, a real-time noise separation and feature extraction module, a fault rule chain deduction module, a multi-dimensional fault judgment module and an intelligent operation and maintenance feedback module; the multi-dimensional sensor array deployment module deploys 9 MEMS micro-vibration sensors at the blade surface, the blade root, the middle of the blade, the blade tip, the contact points of the inner and outer rings of the bearing, the main shaft section, the coupling joint and the bearing support to synchronously collect sub-micron level vibration signals; the noise feature topology library management module constructs a feature library containing all marine environmental noise modes, and each mode stores its topology feature parameters; the real-time noise separation and feature extraction module maps the real-time collected vibration data to a phase space to construct a topology network, separates the noise by calculating the Hamming distance between the vibration data and the modes in the noise library, and extracts a fault feature vector; the fault rule chain deduction module establishes a vibration feature-fault type logical rule chain based on a historical fault case library, which comprises a feature screening, an environmental parameter filtering and a three-layer reasoning mechanism of logical operation; the multi-dimensional fault judgment module inputs the extracted feature vector into the rule chain for deduction, matches through multi-level logical expressions, and outputs the fault type, position and severity; and the intelligent operation and maintenance feedback module generates a structured maintenance report according to the fault judgment result, which comprises a fault position three-dimensional label, a maintenance priority sorting and a historical similar case solution reference, and is pushed to the operation and maintenance personnel terminal in real time.

[0008] The multi-dimensional sensor array deployment module comprises a sub-micron level sensor deployment unit and a space-time consistency synchronous acquisition unit. The sub-micron level sensor deployment unit selects a MEMS capacitive acceleration sensor with a range of ±50g and a resolution of 0.1ug, adopts three-dimensional coordinate positioning technology according to the equipment structure dynamics, and deploys the sensor at 9 stress concentration positions, i.e. the blade surface, the blade root, the middle of the blade, the blade tip, the contact points of the inner and outer rings of the bearing, the main shaft section, the coupling joint and the bearing support, to collect sub-micron level vibration signals in real time. The space-time consistency synchronous acquisition unit uses an internal laser tracker to collect the three-dimensional coordinates of the physical installation positions of the 9 sensors and the structural nodes of the sensor installation positions, and establishes a mapping table of sensor number-three-dimensional coordinates-structural nodes; for the vibration data synchronously collected by the sensors, the error of the time stamps of each channel is corrected through the Kalman filtering algorithm, and the multi-channel data is aligned to a unified time axis with the Beidou satellite clock as the reference; then, the vibration data at each time point is associated with the three-dimensional coordinates of the corresponding sensor according to the mapping table, and a structured data record containing the time stamp, the three-dimensional coordinates and the vibration amplitude is generated, all records are arranged in chronological order, a multi-dimensional vibration data set is formed in HDF5 format, and each data block contains the time dimension, the space dimension and the vibration amplitude dimension, which is convenient for the real-time noise separation and feature extraction module to call.

[0009] The noise feature topology library management module comprises a historical data import unit, a phase space topology feature reconstruction unit and a noise mode clustering modeling unit. The historical data import unit extracts all historical vibration data of marine environment noise from a device full life cycle fault case library, and the data source needs to be labeled with power plant name, collection time and environmental parameters. Then, the unstructured data is converted into a standardized format by an ETL tool: each record contains a time series, environmental parameters and a noise type label. The phase space topology feature reconstruction unit reconstructs the five-dimensional phase space based on the Takens embedding theorem for the historical imported noise time domain data. Specifically, the noise time domain sequence x(t) is taken as the input, and the five-dimensional phase space vector is generated according to the formula X(t) = [x(t), x(t+T), x(t+2T), x(t+3T), x(t+4T)], wherein the embedding dimension m is determined by the false neighbor method, and the time delay T is determined by the first minimum value of the mutual information function. After reconstruction, the Delaunay triangulation algorithm is used to connect the topology of the phase space point set, and through the empty circle property, no other nodes are contained in the incircle of any triangle, forming a topology network reflecting the data distribution density. Finally, three types of topology indicators of the topology network are calculated in real time, namely node connectivity, shortest path length and betweenness centrality. For node connectivity, the number of connections of each node with adjacent nodes is counted, and the average number of connections of all network nodes is taken as the node connectivity. For the shortest path length, the shortest path distance between all node pairs is calculated, and the global shortest path length is taken as the average of the whole network. For the betweenness centrality, the proportion of the number of times each node appears in the shortest path of all node pairs is counted, reflecting the key degree of the node in network information transmission. Then, the average and variance of the three types of indicators are combined into a 10-dimensional feature vector, providing a structured feature input for noise mode clustering. The noise mode clustering modeling unit uses a density clustering algorithm to perform unsupervised classification on the 10-dimensional feature vector, divides it into same-class feature vectors and different-class feature vectors, and clusters the same-class feature vectors and distinguishes the different-class feature vectors. Then, the range intervals of the topology features of various noise modes generated by clustering are established, and the topology feature ranges of various classes are associated with corresponding marine environment types one by one and stored in a data structure, forming a noise feature library to provide a benchmark for real-time noise separation.

[0010] The real-time noise separation and feature extraction module comprises a topology mapping noise separation unit and a fault feature vector extraction unit. The topology mapping noise separation unit maps the real-time vibration data output by the multi-dimensional sensor array deployment module to a five-dimensional phase space based on the Takens embedding theorem and the embedding dimension and time delay parameters pre-trained by the noise feature topology library management module, and then constructs a real-time topology network using the Delaunay triangulation algorithm; then, the Hamming distance between the network and each mode feature vector in the noise feature library is calculated, and when the distance exceeds a preset threshold, it is determined as noise data, and noise separation is realized through a topology edge weight screening mechanism, and the effective vibration data after noise removal is output to the fault feature vector extraction unit; wherein the Hamming distance refers to the standardized count of the difference in the corresponding dimension topology index; the calculation logic is as follows: let the feature vector of the real-time topology network be X = [x1, x2,,, x n ], and the feature vector of a mode in the noise library be Y = [y1, y2,,, y n ], wherein n is the feature dimension; the Hamming distance calculation formula is as follows: ; Wherein, δ(x i , y i ) is an indicator function, which is 1 when |x i −y i |>ε i , and 0 otherwise; ε i is the difference threshold of the i-th feature; The Hamming distance reflects the quantitative comparison of topology features through the standardized dimension difference count, and provides a mathematical basis for noise separation; The fault feature vector extraction unit reconstructs the five-dimensional phase space topology network based on the three-dimensional CAD model of the device and the structural dynamics finite element analysis results for the denoised vibration data output by the topology mapping noise separation unit; then, the high-weight path is identified through the following steps: first, assign a weight to each connected edge according to the node stress distribution obtained from the finite element analysis, the weight value = node stress amplitude / global stress average, calculate the energy transfer efficiency of each path based on the vibration energy flow theory, the efficiency = path energy flux / total energy flux, set an energy proportion threshold, and select the connected subgraph and high-weight path; The morphological algorithm is further used to quantize 6 geometric parameters: path curvature, node betweenness, connected subgraph density, topological network diameter, clustering coefficient and edge weight variance; wherein, the path curvature is obtained by calculating the root mean square deviation of the path node coordinates from the ideal straight line; the node betweenness is obtained by counting the participation frequency of the node in all shortest paths; the connected subgraph density is obtained by calculating the ratio of the actual number of edges of the subgraph to the number of completely connected edges; the topological network diameter is the length of the longest shortest path in the network; the clustering coefficient is obtained by calculating the ratio of the number of connected edges between the neighbors of the node to the maximum possible number of connected edges; and the edge weight variance is obtained by calculating the standard deviation of all connected edge weights; Meanwhile, 4 types of dynamic parameters are quantified: vibration amplitude distribution entropy, characteristic frequency component proportion, time-frequency domain cross-correlation coefficient and energy distribution gradient; Among them, the vibration amplitude distribution entropy is calculated based on the amplitude probability density function of the denoised vibration signal; the characteristic frequency component proportion is determined by the normalized value of the sub-hertz frequency spectrum energy; the time-frequency domain cross-correlation coefficient refers to the correlation index of the time domain signal and the frequency domain feature; and the energy distribution gradient is the vibration energy attenuation rate along the high weight path; Finally, the above 10 types of parameters are standardized and combined into a feature vector, which is complementary to the noise feature output by the noise feature topology library management module, and is directly input into the fault rule chain deduction module to provide quantitative basis for fault type judgment by fusing structural geometric features and dynamic characteristics.

[0011] The fault rule chain deduction module comprises a rule library construction unit and a multi-layer reasoning execution unit. The rule library construction optimization unit extracts more than 1000 groups of fault data verified by actual maintenance from the historical fault case library, and establishes a ternary mapping relationship of feature vector-fault type-environment condition through structured processing; based on the principles of structural mechanics and fault mechanism analysis, the fault feature vector is decomposed into geometric features and dynamic features, and combined with marine environmental parameters to construct a three-layer rule chain: Feature screening layer: set abnormal threshold of each dimension feature; Environment filtering layer: establish environment parameter-feature effectiveness association table; Logic operation layer: combine feature conditions using AND / OR / NOT logic gates; Among them, the feature effectiveness refers to the quantitative correlation degree between each dimension feature value in the fault feature vector and the actual fault type of the equipment under the condition of a specific marine environmental parameter; The multi-layer reasoning execution unit inputs the 10-dimensional feature vector output by the fault feature vector extraction unit and the real-time environmental parameters collected by the device sensor into a three-layer rule chain for reasoning: first, the feature screening layer compares each dimension parameter in the feature vector with the pre-stored threshold condition in the rule base, and marks the abnormal feature dimension that exceeds the preset threshold; Then, the environmental filtering layer calls the environmental parameter-feature threshold correction association rule based on the current marine environmental parameters, which is established based on the mapping relationship between environmental parameters and feature threshold values in historical fault cases, adjusts the threshold condition of the feature screening layer, specifically, according to the real-time environmental parameter interval, retrieves the corresponding feature threshold correction coefficient from the association rule base, and then calculates the preset threshold of the feature screening layer and the correction coefficient according to the formula default threshold × correction coefficient to generate a corrected threshold that adapts to the real-time environment; Finally, the logic operation layer uses a forward reasoning algorithm to logically combine the marked abnormal features and the threshold condition corrected by the environment, matches the logic expression of the feature combination-fault type in the rule base, and deduces the fault type, occurrence location and severity level; During the reasoning process, the noise separation result of the noise feature topology library is called in real time to exclude environmental noise interference.

[0012] The multi-dimensional fault determination module includes a comprehensive reasoning verification unit and a confidence assessment unit; The comprehensive reasoning verification unit performs multi-source fusion verification on the preliminary fault result output by the fault rule chain reasoning module and the denoising data of the real-time noise separation module, the device operating parameters and the environmental parameters; Specifically: first, calculate the dimensional difference between the fault feature vector and the standard feature in the rule base by Euclidean distance; Then, call the running parameter-fault sensitivity association rule in the rule base to verify whether the current speed and load of the device trigger the fault sensitivity condition; Finally, use the sensor number-three-dimensional coordinate-structure node mapping table established by the spatiotemporal consistency synchronous acquisition unit to calculate the spatial distance between the fault location coordinates and the sensor collection points by the Euclidean distance formula, and when the error is less than the preset threshold, determine that the spatial correlation is valid; Through the three-layer verification mechanism, false positives caused by environmental interference are eliminated, and the final fault determination result of the fused structure features, operating state and spatial position is output; The confidence assessment unit quantifies the fault determination result output by the comprehensive reasoning verification unit in multiple dimensions: Rule matching degree calculation: based on the number of rules matched by the logic operation layer and the rule weight given by the rule base construction unit through more than 1000 historical cases, a basic confidence coefficient is generated; Historical similarity retrieval: query the number of cases with the same feature combination in the historical fault case library, and generate a similarity score according to the consistency of the case repair results; The environmental robustness evaluation: based on the mapping relationship between the environmental parameters and the effectiveness of the features in the historical cases, the environmental parameter-confidence correction association rules in the rule library are called to correct the confidence according to the current environmental parameters; when the comprehensive confidence is less than 80%, the secondary reasoning mechanism is triggered: the fault feature vector extraction process is traced back, the noise separation result is checked based on the noise feature topology library, the noise matching degree is calculated by calling the Hamming distance, and the effectiveness of the 10-dimensional feature vector is verified to form a closed-loop optimization of the determination result.

[0013] The intelligent operation and maintenance feedback module includes a structured report generation unit and a real-time terminal pushing unit. The structured report generation unit generates a three-dimensional visualization label of the fault location based on the fault type, location, severity and confidence result output by the multi-dimensional fault determination module, calls the device three-dimensional CAD model to generate the three-dimensional visualization label of the fault location, and distinguishes the fault severity by color coding; the maintenance priority coefficient is calculated in combination with the device operating parameters and the marine environmental parameters; meanwhile, the cases with the same feature combination in the historical fault case library are searched, the solutions, maintenance time consumption and spare parts list are extracted, a structured maintenance report containing the three-dimensional label diagram, maintenance priority ranking and historical case reference is generated, and the report format adopts the XML structure conforming to the marine engineering standard; The real-time terminal pushing unit establishes an encrypted transmission channel by using the Beidou satellite communication module, pushes the structured maintenance report to the operation and maintenance personnel terminal in real time, and the pushing content includes the fault early warning pop-up window, the maintenance report attachment and the real-time communication interface; the system automatically records the pushing time, the receiving state and the operation and maintenance personnel confirmation information to form a closed-loop operation and maintenance record chain.

[0014] A marine engineering data management method based on machine learning, comprising the following steps: S1, deploying sensors at the parts of the device prone to stress concentration, synchronously collecting sub-micron level vibration data and generating a multi-dimensional data set containing time, space and amplitude dimensions through space-time consistency processing; S2, extracting marine environmental noise data from the historical fault case library, clustering to generate a noise feature library through phase space reconstruction and topological feature calculation, and providing a comparison benchmark for real-time noise separation; S3, mapping the real-time vibration data to the phase space to construct a topological network, separating the noise by comparison with the noise feature library, and extracting a fault feature vector containing geometric and dynamic characteristics based on the device structure model; S4, constructing a rule chain based on historical fault cases, inputting the extracted feature vector and real-time environmental parameters into the rule chain to deduce the fault type, location and severity through multi-layer logic matching; S5, further verifying the deduced result through multi-source fusion, calculating the rule matching degree, historical similarity and environmental robustness confidence, and triggering secondary reasoning optimization to determine the result when the threshold is lower. S6. Generate a structured report containing 3D annotations, maintenance priorities, and historical cases based on the final fault determination. The report is then encrypted and pushed to the maintenance terminal via BeiDou communication to form a closed-loop maintenance record.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Improved noise separation accuracy: A noise feature library (containing 1,000 sets of historical noise data) is constructed using phase space reconstruction technology. Combined with the Hamming distance matching algorithm, the residual noise energy ratio after noise separation is less than 5%, which significantly improves the reliability of fault feature identification.

[0016] 2. Breakthrough in 3D spatial positioning accuracy: Based on the equipment's 3D CAD model (accuracy 0.1mm) and sensor array mapping table (positioning accuracy 0.05mm), the fault location error is ≤0.1mm through Euclidean distance calculation, meeting the precision maintenance needs of deep-sea equipment.

[0017] 3. Significantly reduced misjudgment rate: The confidence assessment unit uses three-dimensional verification of rule matching degree (weight coefficient), historical similarity (matching cases ≥30 groups) and environmental robustness (correction coefficient 0.9) to reduce the overall misjudgment rate from 30% of the traditional method to below 8%. Attached Figure Description

[0018] Figure 1 This is a system workflow diagram of a machine learning-based marine engineering data management system according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example: Figure 1 As shown, the present invention provides a technical solution. The marine engineering data management system based on machine learning comprises a multi-dimensional sensor array deployment module, a noise feature topology library management module, a real-time noise separation and feature extraction module, a fault rule chain deduction module, a multi-dimensional fault judgment module and an intelligent operation and maintenance feedback module; the multi-dimensional sensor array deployment module deploys 9 MEMS micro-vibration sensors at the blade surface, the blade root, the middle of the blade, the blade tip, the contact points of the inner and outer rings of the bearing, the main shaft section, the coupling joint and the bearing support to synchronously collect sub-micron level vibration signals; the noise feature topology library management module constructs a feature library containing all marine environmental noise modes, and each mode stores its topology feature parameters; the real-time noise separation and feature extraction module maps the real-time collected vibration data to a phase space to construct a topology network, separates the noise by calculating the Hamming distance between the vibration data and the modes in the noise library, and extracts a fault feature vector; the fault rule chain deduction module establishes a vibration feature-fault type logical rule chain based on a historical fault case library, which comprises a feature screening, an environmental parameter filtering and a three-layer reasoning mechanism of logical operation; the multi-dimensional fault judgment module inputs the extracted feature vector into the rule chain for deduction, matches through multi-level logical expressions, and outputs the fault type, position and severity; and the intelligent operation and maintenance feedback module generates a structured maintenance report according to the fault judgment result, which comprises a fault position three-dimensional label, a maintenance priority sorting and a historical similar case solution reference, and is pushed to the operation and maintenance personnel terminal in real time.

[0021] The multi-dimensional sensor array deployment module comprises a sub-micron level sensor deployment unit and a space-time consistency synchronous acquisition unit. The sub-micron level sensor deployment unit selects a MEMS capacitive acceleration sensor with a range of ±50g and a resolution of 0.1ug, adopts three-dimensional coordinate positioning technology according to the equipment structure dynamics, and deploys the sensor at 9 stress concentration parts, i.e. the blade surface, the blade root, the middle of the blade, the blade tip, the contact points of the inner and outer rings of the bearing, the main shaft section, the coupling joint and the bearing support, to collect sub-micron level vibration signals in real time. The space-time consistency synchronous acquisition unit uses an internal laser tracker to collect the three-dimensional coordinates of the physical installation positions of the 9 sensors and the structural nodes of the sensor installation positions, and establishes a mapping table of sensor number-three-dimensional coordinates-structural nodes; for the vibration data synchronously collected by the sensors, the error of the time stamps of each channel is corrected through the Kalman filtering algorithm, and the multi-channel data is aligned to a unified time axis with the Beidou satellite clock as the reference; then, the vibration data at each time point is associated with the three-dimensional coordinates of the corresponding sensor according to the mapping table, and a structured data record containing the time stamp, the three-dimensional coordinates and the vibration amplitude is generated, all records are arranged in chronological order, a multi-dimensional vibration data set is formed in HDF5 format, and each data block contains the time dimension, the space dimension and the vibration amplitude dimension, which is convenient for the real-time noise separation and feature extraction module to call.

[0022] The noise feature topology library management module comprises a historical data import unit, a phase space topology feature reconstruction unit and a noise mode clustering modeling unit. The historical data import unit extracts all historical vibration data of marine environment noise from a device full life cycle fault case library, and the data source needs to be labeled with power plant name, collection time and environmental parameters, and then non-structured data is converted into a standardized format by an ETL tool: each record contains a time series, environmental parameters and noise type label. The phase space topology feature reconstruction unit reconstructs the five-dimensional phase space based on the Takens embedding theorem, and the specific steps are as follows: taking the noise time domain sequence x(t) as input, generating a five-dimensional phase space vector according to the formula X(t) = [x(t), x(t+T), x(t+2T), x(t+3T), x(t+4T)], wherein the embedding dimension m is determined by the false neighbor method, and the time delay T is determined by the first minimum value of the mutual information function; after reconstruction, the Delaunay triangulation algorithm is used to connect the topology of the phase space point set, and through the empty circle property, no other nodes are contained in the incircle of any triangle, forming a topology network reflecting the data distribution density; finally, three types of topology indexes of the topology network are calculated in real time, namely node connectivity, shortest path length and betweenness centrality; for node connectivity, the number of connections of each node with adjacent nodes is counted, and the average number of connections of all network nodes is taken as the node connectivity; for the shortest path length, the shortest path distance between all node pairs is calculated, and the global shortest path length is taken as the average value of the whole network; for the betweenness centrality, the proportion of the number of times each node appears in the shortest path of all node pairs is counted, reflecting the key degree of the node in network information transmission; then the average and variance of the three types of indexes are combined into a 10-dimensional feature vector, providing structured feature input for noise mode clustering; The noise mode clustering modeling unit uses a density clustering algorithm to perform unsupervised classification on the 10-dimensional feature vector, divides it into same-class feature vectors and different-class feature vectors, and clusters the same-class feature vectors and distinguishes the different-class feature vectors; then, for various noise modes generated by clustering, the range interval of the topology feature is established, and each type of topology feature range is associated with the corresponding marine environment type one by one and stored in a data structure, forming a noise feature library to provide a benchmark for real-time noise separation.

[0023] The real-time noise separation and feature extraction module comprises a topology mapping noise separation unit and a fault feature vector extraction unit. The topology mapping noise separation unit maps the real-time vibration data output by the multi-dimensional sensing array deployment module to a five-dimensional phase space based on the Takens embedding theorem and using the embedding dimension and time delay parameters pre-trained by the noise feature topology library management module, and then constructs a real-time topology network using a Delaunay triangulation algorithm; then, the Hamming distance between the network and each mode feature vector in the noise feature library is calculated, and when the distance exceeds a preset threshold, the data is determined to be noise data, and noise separation is achieved through a topology edge weight screening mechanism, and the effective vibration data after noise removal is output to the fault feature vector extraction unit; wherein the Hamming distance refers to the standardized count of the difference in the corresponding dimension topology index; the calculation logic is as follows: let the feature vector of the real-time topology network be X = [x1, x2,,, x n ], and the feature vector of a mode in the noise library be Y = [y1, y2,,, y n ], wherein n is the feature dimension; the Hamming distance calculation formula is as follows: ; Wherein, δ(x i , y i ) is an indicator function, which is 1 when |x i −y i |>ε i , and 0 otherwise; ε i is the difference threshold of the i-th feature; The Hamming distance reflects the quantitative comparison of topology features through the standardized dimension difference count, and provides a mathematical basis for noise separation; The fault feature vector extraction unit reconstructs the five-dimensional phase space topology network based on the three-dimensional CAD model of the device and the structural dynamics finite element analysis results for the denoised vibration data output by the topology mapping noise separation unit; then, the high-weight path is identified through the following steps: first, each connected edge is given a weight according to the node stress distribution obtained from the finite element analysis, and the weight value = node stress amplitude / global stress average; based on the vibration energy flow theory, the energy transfer efficiency of each path is calculated, efficiency = path energy flux / total energy flux; set an energy proportion threshold to screen out connected subgraphs and high-weight paths; The morphological algorithm is further used to quantize six types of geometric parameters: path curvature, node betweenness, connected subgraph density, topological network diameter, clustering coefficient and edge weight variance. The path curvature is obtained by calculating the root mean square deviation of the path node coordinates from an ideal straight line. The node betweenness is obtained by counting the frequency of participation of the node in all shortest paths. The connected subgraph density is obtained by calculating the ratio of the actual number of edges of the subgraph to the number of completely connected edges. The topological network diameter is the length of the longest shortest path in the network. The clustering coefficient is obtained by calculating the ratio of the number of connecting edges between the neighbors of the node to the maximum possible number of connecting edges. The edge weight variance is obtained by calculating the standard deviation of the weight of all connecting edges. Meanwhile, four types of dynamic parameters are quantified: vibration amplitude distribution entropy, characteristic frequency component proportion, time-frequency domain cross-correlation coefficient and energy distribution gradient. The vibration amplitude distribution entropy is calculated based on the amplitude probability density function of the denoised vibration signal. The characteristic frequency component proportion is determined by the normalized value of the frequency spectrum energy in the sub-hertz frequency band. The time-frequency domain cross-correlation coefficient refers to the correlation index of the time domain signal and the frequency domain feature. The energy distribution gradient is the vibration energy attenuation rate along the high weight path. Finally, the above 10 types of parameters are standardized and combined into a feature vector, which is complementary to the noise feature output by the noise feature topology library management module, and is directly input into the fault rule chain deduction module to provide quantitative basis for fault type judgment by fusing structural geometric features and dynamic characteristics.

[0024] The fault rule chain deduction module includes a rule library construction unit and a multi-layer reasoning execution unit. The rule library construction optimization unit extracts more than 1000 groups of fault data verified by actual maintenance from the historical fault case library, and establishes a ternary mapping relationship of feature vector-fault type-environment condition through structured processing. Based on the principles of structural mechanics and fault mechanism analysis, the fault feature vector is decomposed into geometric features and dynamic features, and combined with marine environmental parameters to construct a three-layer rule chain: Feature screening layer: set abnormal threshold of each dimension feature; Environment filtering layer: establish environment parameter-feature effectiveness association table; Logic operation layer: combine feature conditions using AND / OR / NOT logic gates; Among them, the feature effectiveness refers to the quantitative correlation degree between each dimension feature value in the fault feature vector and the actual fault type of the equipment under the condition of a specific marine environmental parameter; The multi-layer reasoning execution unit inputs the 10-dimensional feature vector output by the fault feature vector extraction unit and the real-time environmental parameters collected by the device sensor into a three-layer rule chain for reasoning: first, the feature screening layer compares each dimension parameter in the feature vector with the pre-stored threshold condition in the rule base, and marks the abnormal feature dimension that exceeds the preset threshold; Then, the environmental filtering layer calls the environmental parameter-feature threshold correction association rule based on the current marine environmental parameters, which is established based on the mapping relationship between environmental parameters and feature threshold values in historical fault cases, adjusts the threshold condition of the feature screening layer, specifically, according to the real-time environmental parameter interval, retrieves the corresponding feature threshold correction coefficient from the association rule base, and then calculates the preset threshold of the feature screening layer and the correction coefficient according to the formula default threshold × correction coefficient to generate a corrected threshold that adapts to the real-time environment; Finally, the logic operation layer uses a forward reasoning algorithm to logically combine the marked abnormal features and the threshold condition corrected by the environment, matches the logic expression of the feature combination-fault type in the rule base, and deduces the fault type, occurrence location and severity level; During the reasoning process, the noise separation result of the noise feature topology library is called in real time to exclude environmental noise interference.

[0025] The multi-dimensional fault judgment module includes a comprehensive reasoning verification unit and a confidence assessment unit; The comprehensive reasoning verification unit performs multi-source fusion verification on the preliminary fault result output by the fault rule chain reasoning module and the denoising data of the real-time noise separation module, the device operating parameters and the environmental parameters; Specifically: first, calculate the dimensional difference between the fault feature vector and the standard feature in the rule base by Euclidean distance; Then, call the running parameter-fault sensitivity association rule in the rule base to verify whether the current speed and load of the device trigger the fault sensitivity condition; Finally, use the sensor number-three-dimensional coordinate-structure node mapping table established by the space-time consistency synchronous acquisition unit to calculate the spatial distance between the fault location coordinates and the sensor collection points by the Euclidean distance formula, and when the error is less than the preset threshold, determine that the spatial correlation is valid; Through the three-layer verification mechanism, false positives caused by environmental interference are eliminated, and the final fault judgment result of the fused structure features, operating state and spatial position is output; The confidence assessment unit quantifies the fault judgment result output by the comprehensive reasoning verification unit in multiple dimensions: Rule matching degree calculation: based on the number of rules matched by the logic operation layer and the rule weight given by the rule base construction unit through more than 1000 historical cases, a basic confidence coefficient is generated; Historical similarity retrieval: query the number of cases with the same feature combination in the historical fault case library, and generate a similarity score according to the consistency of the case repair results; The environmental robustness evaluation: based on the mapping relationship between the environmental parameters and the effectiveness of the features in the historical cases, the environmental parameter-confidence correction association rules in the rule library are called to correct the confidence according to the current environmental parameters; when the comprehensive confidence is less than 80%, the secondary reasoning mechanism is triggered: the fault feature vector extraction process is traced back, the noise separation result is checked based on the noise feature topology library, the noise matching degree is calculated by calling the Hamming distance, and the effectiveness of the 10-dimensional feature vector is verified to form a closed-loop optimization of the judgment result.

[0026] The intelligent operation and maintenance feedback module includes a structured report generation unit and a real-time terminal pushing unit. The structured report generation unit generates a three-dimensional visualization label of the fault location by calling a device three-dimensional CAD model based on the fault type, location, severity and confidence result output by the multi-dimensional fault judgment module, and distinguishes the fault severity by color coding; the maintenance priority coefficient is calculated in combination with the device operation parameters and the marine environmental parameters; meanwhile, the cases with the same feature combination in the historical fault case library are searched, the solutions, maintenance time consumption and spare parts list are extracted, a structured maintenance report containing the three-dimensional label diagram, the maintenance priority ranking and the historical case reference is generated, and the report format adopts the XML structure conforming to the marine engineering standard; The real-time terminal pushing unit establishes an encrypted transmission channel by using the Beidou satellite communication module, and pushes the structured maintenance report to the operation and maintenance personnel terminal in real time, and the pushing content includes the fault early warning pop-up window, the maintenance report attachment and the real-time communication interface; the system automatically records the pushing time, the receiving state and the operation and maintenance personnel confirmation information to form a closed-loop operation and maintenance record chain.

[0027] A marine engineering data management method based on machine learning, comprising the following steps: S1, deploying sensors at the stress concentration parts of the equipment, synchronously collecting sub-micron level vibration data and generating a multi-dimensional data set containing time, space and amplitude dimensions through space-time consistency processing; S2, extracting marine environmental noise data from the historical fault case library, clustering to generate a noise feature library through phase space reconstruction and topological feature calculation, and providing a comparison benchmark for real-time noise separation; S3, mapping the real-time vibration data to the phase space to construct a topological network, separating the noise by comparison with the noise feature library, and extracting a fault feature vector containing geometric and dynamic characteristics based on the equipment structure model; S4, constructing a rule chain based on historical fault cases, inputting the extracted feature vector and real-time environmental parameters into the rule chain to deduce the fault type, location and severity through multi-layer logic matching; S5, further verifying the deduced result through multi-source fusion, calculating the rule matching degree, the historical similarity and the environmental robustness confidence, and triggering the secondary reasoning optimization judgment result when the threshold is lower. S6、According to the final fault judgment, a structured report containing three-dimensional labels, maintenance priority and historical cases is generated, and is pushed to the operation and maintenance terminal through Beidou communication encryption to form a closed-loop operation and maintenance record.

[0028] On a 3MW wind turbine of a certain offshore wind power platform in the South China Sea, according to the results of equipment structure dynamics analysis, MEMS capacitive acceleration sensors with a range of ±50g and a resolution of 0.1μg are deployed at 9 stress concentration positions on the blade surface (1 / 3 away from the blade root), the blade root, the middle of the blade, the blade tip, the inner and outer contact points of the bearing, the main shaft section (500mm away from the bearing), the coupling joint and the bearing support. A laser tracker (accuracy 0.05mm) is used to collect the three-dimensional coordinates of each sensor, and a mapping table is established, such as: sensor S1 corresponds to the blade root coordinates (X=12.5m, Y=3.2m, Z=8.7m), and the structure node is the blade-hub connecting tenon.

[0029] The sensors are synchronized by a space-time consistency synchronization acquisition unit to synchronously acquire vibration data at a sampling rate of 1024Hz. For example, at a certain time t=10:23:45.678, the vibration amplitude acquired by sensor S1 is 12.5μg, the time stamp error is corrected by Kalman filtering (the time stamp error after correction is less than or equal to 1ms from the Beidou clock), and the three-dimensional coordinates are associated to generate a structured record: [t=10:23:45.678, X=12.5m, Y=3.2m, Z=8.7m, amplitude=12.5μg]. All data are stored in HDF5 format to form a three-dimensional data set containing time (10000 time points), space (9 sensor coordinates) and amplitude (-50g~50g) for subsequent module calling.

[0030] 1000 sets of marine environmental noise data are extracted from the platform's full life cycle case library, such as typhoon period (sea current speed 3.5m / s), high tide (wave height 2.8m) and other scene vibration data. Taking a set of typhoon period noise time domain sequence x(t) as an example, five-dimensional phase space reconstruction is carried out based on Takens embedding theorem: The embedding dimension m=5 (the false neighbor rate is less than 5% calculated by the false neighbor method), and the time delay T=10 sampling points (at the first minimum value of mutual information function), the phase space vector X(t)=[x(t),x(t+10),x(t+20),x(t+30),x(t+40)] is generated.

[0031] Delaunay triangulation is performed on 1000 phase space points to construct a topological network. Topological indicators are calculated: the average node connectivity is 4.2 (statistically, each node is connected to an average of 4.2 adjacent nodes), the average global shortest path length is 2.8 (the average distance of the shortest path between all node pairs), and the average betweenness centrality is 0.15 (the average proportion of a node's appearance in the shortest path). The average and variance of the three indicators are combined into a 10-dimensional feature vector, such as: [4.2, 0.5, 2.8, 0.3, 0.15, 0.08,...] (the first three dimensions are the average and variance of node connectivity, the middle three dimensions are the average and variance of the shortest path length, and the last three dimensions are the average and variance of betweenness centrality).

[0032] The 10-dimensional feature vector is classified using a density clustering algorithm, such as grouping typhoon noise into a class with a node connectivity range of [3.8, 4.6] and a shortest path length range of [2.5, 3.1]. A "topological feature range - marine environment type" correlation table is established, such as node connectivity > 4.0 and shortest path length < 3.0 corresponding to typhoon noise mode, forming a noise feature library stored in a distributed database.

[0033] On July 1, 2025, at 14:00, the system collected real-time vibration data, and based on the noise feature library pre-trained with embedding dimension m=5 and time delay T=10, it was mapped to a five-dimensional phase space and constructed a real-time topological network. The Hamming distance between this network and the typhoon noise mode in the noise library is calculated: The real-time feature vector X=[4.3, 0.6, 2.7, 0.25, 0.16,...], the noise library feature vector Y=[4.0, 0.5, 2.8, 0.3, 0.15,...], and the threshold value εi for each dimension is set to 0.2 (such as the node connectivity threshold 0.2).

[0034] Calculate δ(xi,yi): when |4.3-4.0|=0.3>0.2, δ=1; |0.6-0.5|=0.1≤0.2, δ=0, and so on. The Hamming distance DH=(1+0+1+0+0+...) / 10=0.3 (assuming 3 dimensions exceed the threshold), which exceeds the preset threshold value 0.25, and is determined as typhoon noise. Through topological edge weight screening (retaining edges with weight > 0.7), the noise is separated, and the denoised data is output.

[0035] For the denoised data, the topology network is reconstructed based on the 3D CAD model of the device (mesh accuracy 0.1mm). According to the finite element analysis, the stress amplitude of a certain path node is 150MPa, the global stress average is 100MPa, and the edge weight = 150 / 100 = 1.5; the path energy flux is 2.3W, the total energy flux is 3.0W, and the energy transfer efficiency = 2.3 / 3.0 ≈ 76.7%, which exceeds the 70% threshold, and is determined as a high-weight path. Quantify the geometric parameters: Path curvature: Calculate the root mean square deviation of the path node coordinates from the ideal straight line, assuming that the sum of the squared deviations of 10 node coordinates is 0.01mm², and the root mean square deviation is 0.1mm; Node betweenness: Statistics of this node in 100 shortest paths participate 30 times, betweenness = 30 / 100 = 0.3; Combine with kinetic parameters such as characteristic frequency component proportion (sub-hertz band 0.3Hz energy proportion 25%), generate 10-dimensional fault feature vector: [0.1, 0.3, 0.65, 5.2, 0.4, 0.2, 0.85, 25%, 0.9, 0.15] (the first six dimensions are geometric parameters, and the last four dimensions are kinetic parameters).

[0036] Input the feature vector into the fault rule chain deduction module: Feature screening layer: preset path curvature threshold 0.1mm, node betweenness threshold 0.25, current value 0.1mm (not exceeding threshold), 0.3 (exceeding threshold), mark node betweenness as abnormal dimension.

[0037] Environmental filtering layer: real-time sea current speed 2.2m / s, call associated rule "when sea current speed > 2m / s, node betweenness threshold floats up by 20%", modify threshold to 0.25x1.2 = 0.3, current value 0.3 does not exceed modified threshold, cancel abnormal mark.

[0038] Logical operation layer: combine with other features (such as 0.3Hz energy proportion 25% > 20%), match logical expression "(node betweenness > 0.25 ∧ 0.3Hz energy proportion > 20%) → blade tenon loose", deduce fault type as blade tenon loose, location as blade root, severity as medium.

[0039] Multi-dimensional fault judgment module verification: Calculate the Euclidean distance between the feature vector and the standard feature vector: assume the standard vector is [0.1, 0.25, 0.6, 5.0, 0.4, 0.2, 0.8, 20%, 0.9, 0.1], Euclidean distance =√[(0.1-0.1)²+(0.3-0.25)²+...]=0.07, less than threshold 0.1, feature consistency is valid; Spatial correlation: the distance between the fault location blade root coordinate and the sensor S1 coordinate is 0.08m < 0.1mm, which is determined to be valid; Confidence evaluation: rule matching degree 0.8 (matching 2 key rules), historical similarity 0.9 (35 similar cases), environmental robustness 0.85 (current speed correction coefficient 0.85), comprehensive confidence = 0.8x0.3+0.9x0.4+0.85x0.3=0.86>0.8, no secondary reasoning is needed.

[0040] According to the determination result, the structured report generation unit calls the three-dimensional CAD model, generates a red three-dimensional label (accuracy 0.1mm) at the blade root position, calculates the maintenance priority coefficient: severity medium (coefficient 2) x environmental influence factor (current speed 2.2m / s corresponds to factor 1.5) = 3, priority level 3 (1-5 levels). Retrieve the historical case library, extract the average maintenance time of 35 groups of loose blade tenon cases (4 hours), spare parts list (screw x2, torque wrench), generate an XML format report; The real-time terminal pushing unit converts the report into JSON format and pushes it to the operation and maintenance personnel tablet through the Beidou satellite communication module, and prompts the "blade root tenon loose, suggest repairing within 4 hours" pop-up window. After receiving, the operation and maintenance personnel confirm the processing, the system records the pushing time 14:15 and the receiving state has been read, forms the operation and maintenance record chain, and completes the whole process closed loop from data collection to maintenance decision.

[0041] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, and that the present application can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, the scope of the present application being defined by the appended claims rather than the foregoing description, and it is intended that all changes that come within the meaning and range of equivalency of the claims are embraced therein. Any reference signs in the claims should not be construed as limiting the claims to the figures in which the reference signs are used.

Claims

1. A marine engineering data management system based on machine learning, characterized in that: The system includes a multi-dimensional sensor array deployment module, a noise feature topology library management module, a real-time noise separation and feature extraction module, a fault rule chain deduction module, a multi-dimensional fault judgment module, and an intelligent operation and maintenance feedback module. The multi-dimensional sensor array deployment module deploys nine MEMS micro-vibration sensors on the blade surface, blade root, blade middle, blade tip, bearing inner and outer ring contact points, main shaft section, coupling joint, and bearing support to synchronously acquire sub-micron level vibration signals. The noise feature topology library management module constructs a feature library containing all marine environmental noise patterns, storing the topological feature parameters for each pattern. The real-time noise separation and feature extraction module maps the real-time acquired vibration data to a phase space structure. A topology network is constructed, and noise is separated by calculating its Hamming distance with patterns in the noise database, thereby extracting fault feature vectors. The fault rule chain inference module establishes a logical rule chain of vibration features and fault types based on a historical fault case database, including a three-layer inference mechanism of feature screening, environmental parameter filtering, and logical operation. The multi-dimensional fault judgment module inputs the extracted feature vectors into the rule chain for inference, and outputs the fault type, location, and severity through multi-level logical expression matching. The intelligent operation and maintenance feedback module generates a structured maintenance report based on the fault judgment results, including three-dimensional annotation of the fault location, maintenance priority ranking, and historical similar case solution references, and pushes it to the operation and maintenance personnel's terminal in real time.

2. The marine engineering data management system based on machine learning according to claim 1, characterized in that: The multi-dimensional sensor array deployment module includes a submicron-level sensor deployment unit and a spatiotemporally consistent synchronous acquisition unit. The submicron-level sensor deployment unit uses a MEMS capacitive accelerometer with a range of ±50g and a resolution of 0.1μg. Based on the equipment's structural dynamics, three-dimensional coordinate positioning technology is used to deploy the sensor at nine locations prone to stress concentration: the blade surface, blade root, blade middle, blade tip, inner and outer bearing contact points, main shaft section, coupling joint, and bearing support. The sensor collects sub-meter-level vibration signals in real time. The spatiotemporal consistency synchronous acquisition unit uses the in-device laser tracker to acquire the three-dimensional coordinates of the physical installation positions of the nine sensors and the structural nodes of the sensor installation positions, and establishes a mapping table of sensor number-three-dimensional coordinates-structural nodes; For vibration data synchronously acquired by sensors, the Kalman filter algorithm is used to correct the error of the timestamps of each channel, aligning the multi-channel data to a unified time axis based on the BeiDou satellite clock. Then, according to the mapping table, the vibration data at each time point is associated with the three-dimensional coordinates of the corresponding sensor to generate a structured data record containing timestamps, three-dimensional coordinates, and vibration amplitude. All records are arranged in chronological order to form a multidimensional vibration dataset stored in HDF5 format. Each data block contains time dimension, spatial dimension, and vibration amplitude dimension, which is convenient for the real-time noise separation and feature extraction module to call.

3. The marine engineering data management system based on machine learning according to claim 1, characterized in that: The noise feature topology library management module includes a historical data import unit, a phase space topology feature reconstruction unit, and a noise pattern clustering modeling unit. The historical data import unit extracts all historical vibration data of marine environmental noise from the equipment life cycle failure case library. The data source must be labeled with the power station name, collection time, and environmental parameters. Then, the unstructured data is converted into a standardized format using an ETL tool: each record contains time series, environmental parameters, and noise type labels. The phase space topology feature reconstruction unit, based on Takens' embedding theorem, performs five-dimensional phase space reconstruction on historically imported noisy time-domain data. Specifically, it takes the noisy time-domain sequence x(t) as input and generates a five-dimensional phase space vector according to the formula X(t) = [x(t), x(t+T), x(t+2T), x(t+3T), x(t+4T)]. The embedding dimension m is determined by the spurious neighbor method, and the time delay T... Sampling points are determined by the first minimum value of the mutual information function. After reconstruction, the Delaunay triangulation algorithm is used to perform topological connections on the phase space point set. Through the empty circle property, the circumcircle of any triangle contains no other nodes, forming a topological network that reflects the data distribution density. Finally, three types of topological indices of the topological network are calculated in real time: node connectivity, shortest path length, and betweenness centrality. For node connectivity, the number of connections between each node and its neighbors is counted, and the average number of connections of all nodes in the network is taken as the node connectivity. For shortest path length, the shortest path distance between all pairs of nodes is calculated, and the average of the shortest path distances of all nodes in the network is taken as the global shortest path length. For betweenness centrality, the proportion of times each node appears in the shortest path of all pairs of nodes is counted, reflecting the criticality of the node in the network information transmission. The mean and variance of the three indices are then combined into a 10-dimensional feature vector to provide structured feature input for noisy pattern clustering. The noise pattern clustering modeling unit uses a density clustering algorithm to perform unsupervised classification of 10-dimensional feature vectors, dividing them into similar feature vectors and dissimilar feature vectors. Similar feature vectors are clustered together, while dissimilar feature vectors are distinguished. Then, for each noise pattern generated by clustering, a range of topological features is established, and each type of topological feature range is associated with its corresponding marine environment type and stored in a data structure to form a noise feature library, providing a benchmark for comparison in real-time noise separation.

4. The marine engineering data management system based on machine learning according to claim 1, characterized in that: The real-time noise separation and feature extraction module includes a topology mapping noise separation unit and a fault feature vector extraction unit; The topology mapping noise separation unit maps the real-time vibration data output by the multi-dimensional sensor array deployment module to a five-dimensional phase space based on Takens' embedding theorem and using the embedding dimension and time delay parameters pre-trained by the noise feature topology library management module. Then, it constructs a real-time topology network using the Delaunay triangulation algorithm. Next, it calculates the Hamming distance between this network and the feature vectors of each mode in the noise feature library. Data exceeding a preset threshold is considered noise data. Noise separation is then achieved through a topology edge weighting filtering mechanism, and the denoised valid vibration data is output to the fault feature vector extraction unit. The Hamming distance refers to the standardized count of differences in topology indicators across corresponding dimensions. The calculation logic is as follows: Let the feature vector of the real-time topology network be X = [x1, x2, ..., x...]. n The feature vector of a certain pattern in the noise database is Y = [y1, y2, ..., y]. n ], where n is the feature dimension; the Hamming distance calculation formula is as follows: ； Where, δ(x) i , y i ) is an indicator function, when |x i -y i ∣>ε i ε is 1 if true, otherwise 0; i The difference threshold for the i-th dimension feature; Hamming distance, through standardized dimensionality difference counting, reflects a quantitative comparison of topological features, providing a mathematical basis for noise separation; The fault feature vector extraction unit reconstructs a five-dimensional phase space topology network based on the denoised vibration data output by the topology mapping noise separation unit, using the equipment's three-dimensional CAD model and the structural dynamics finite element analysis results. Then, high-weight paths are identified through the following steps: First, each connecting edge is assigned a weight based on the node stress distribution obtained from the finite element analysis, with the weight value being the node stress amplitude and the global stress mean. The energy transfer efficiency of each path is calculated based on the vibration energy flow theory, with the efficiency being the path energy flux and the total energy flux. An energy percentage threshold is set to filter out connected subgraphs and high-weight paths. Then, morphological algorithms are used to quantify six types of geometric parameters: path curvature, node betweenness, connected subgraph density, topological network diameter, clustering coefficient, and edge weight variance. Path curvature is obtained by calculating the root mean square deviation of the path node coordinates from the ideal straight line; node betweenness is obtained by statistically analyzing the frequency of node participation in all shortest paths; connected subgraph density is obtained by calculating the ratio of the actual number of edges to the number of fully connected edges; topological network diameter is obtained by using the length of the longest and shortest paths in the network; clustering coefficient is obtained by calculating the ratio of the number of edges connected to a node's neighbors to the maximum possible number of connected edges; and edge weight variance is obtained by calculating the standard deviation of all connected edge weights. Simultaneously, four types of dynamic parameters are quantified: vibration amplitude distribution entropy, proportion of characteristic frequency components, time-frequency domain cross-correlation coefficient, and energy distribution gradient. Among them, the vibration amplitude distribution entropy is calculated based on the amplitude probability density function of the denoised vibration signal; the proportion of characteristic frequency components is determined by the normalized value of the spectral energy in the sub-Hertz band; the time-frequency domain cross-correlation coefficient refers to the correlation index between the time-domain signal and the frequency-domain features; and the energy distribution gradient is the vibration energy decay rate along the high-weight path. Finally, the above 10 types of parameters are standardized and combined into a feature vector. This vector complements the noise features output by the noise feature topology library management module and is directly input into the fault rule chain deduction module to provide a quantitative basis for fault type determination by integrating structural geometric features and dynamic characteristics.

5. A marine engineering data management system based on machine learning according to claim 1, characterized in that: The fault rule chain deduction module includes a rule base construction unit and a multi-layer inference execution unit; The rule base construction and optimization unit extracts over 1000 sets of fault data verified through actual maintenance from the historical fault case database, and establishes a ternary mapping relationship of feature vector-fault type-environmental conditions through structured processing; based on the structural mechanics principles of equipment and fault mechanism analysis, the fault feature vector is decomposed into geometric features and dynamic features, and a three-layer rule chain is constructed by combining marine environmental parameters: Feature filtering layer: Sets the anomaly threshold for features in each dimension; Environmental filtering layer: Establish an environmental parameter-feature validity association table; Logic operation layer: Uses AND / OR / NOT logic gates to combine feature conditions; Among them, feature validity refers to the degree of quantitative correlation between the feature values ​​of each dimension in the fault feature vector and the actual fault type of the equipment under specific marine environmental parameter conditions. The multi-layer inference execution unit inputs the 10-dimensional feature vector output by the fault feature vector extraction unit and the real-time environmental parameters collected by the device sensors into the three-layer rule chain for inference: First, the feature filtering layer compares the parameters of each dimension in the feature vector with the threshold conditions stored in the rule base, and marks the abnormal feature dimensions that exceed the preset threshold. Then, the environmental filtering layer calls the environmental parameter-feature threshold correction association rule based on the current marine environmental parameters. This rule is established based on the mapping relationship between environmental parameters and feature thresholds in historical failure cases. The threshold conditions of the feature filtering layer are adjusted. Specifically, based on the real-time environmental parameter range, the corresponding feature threshold correction coefficient is retrieved from the association rule library. Then, the preset threshold of the feature filtering layer and the correction coefficient are calculated according to the formula default threshold × correction coefficient to generate a correction threshold adapted to the real-time environment. Finally, the forward reasoning algorithm is used in the logic operation layer to logically combine the marked abnormal features with the environmentally corrected threshold conditions, match the logical expression of feature combination-fault type in the rule base, and deduce the fault type, location and severity level. During the inference process, the noise separation results from the noise feature topology library are called in real time to eliminate environmental noise interference.

6. A marine engineering data management system based on machine learning according to claim 1, characterized in that: The multi-dimensional fault determination module includes a comprehensive reasoning verification unit and a confidence evaluation unit; The integrated reasoning verification unit performs multi-source fusion verification of the preliminary fault results output by the fault rule chain inference module and the denoised data, equipment operating parameters, and environmental parameters from the real-time noise separation module. Specifically: First, the dimensional difference between the fault feature vector and the standard features in the rule base is calculated using Euclidean distance; then, the operating parameter-fault sensitivity association rule in the rule base is called to verify whether the current speed and load of the equipment trigger fault-sensitive conditions; finally, the sensor number-3D coordinate-structure node mapping table established by the spatiotemporal consistency synchronous acquisition unit is used to calculate the spatial distance between the fault location coordinates and the sensor acquisition points using the Euclidean distance formula. When the error is less than a preset threshold, the spatial correlation is deemed valid; through a three-layer verification mechanism, misjudgments caused by environmental interference are eliminated, and the final fault judgment result, which integrates structural features, operating status, and spatial location, is output. The confidence evaluation unit performs multi-dimensional confidence quantification on the fault determination results output by the integrated reasoning verification unit: Rule matching degree calculation: Based on the number of rules matched by the logical operation layer and the rule base construction unit, a basic confidence coefficient is generated by assigning rule weights to more than 1,000 historical cases; Historical similarity retrieval: Query the number of cases with the same combination of features in the historical fault case database, and generate a similarity score based on the consistency of the case repair results; Environmental robustness assessment: Based on the mapping relationship between environmental parameters and feature validity in historical cases, the association rules of environmental parameters-confidence in the rule base are called to correct the association rules, and the confidence is corrected according to the current environmental parameters. When the overall confidence is <80%, the secondary inference mechanism is triggered: the fault feature vector extraction process is traced back, the noise separation results are checked based on the noise feature topology library, the Hamming distance is called to calculate the noise matching degree, and the validity of the 10-dimensional feature vector is verified to form a closed-loop optimization of the judgment result.

7. A marine engineering data management system based on machine learning according to claim 1, characterized in that: The intelligent operation and maintenance feedback module includes a structured report generation unit and a real-time terminal push unit; The structured report generation unit, based on the fault type, location, severity, and confidence results output by the multi-dimensional fault judgment module, calls the equipment's 3D CAD model to generate a 3D visual annotation of the fault location, and distinguishes the severity of the fault through color coding; it calculates the maintenance priority coefficient by combining equipment operating parameters and marine environmental parameters; at the same time, it retrieves cases with the same combination of features from the historical fault case library, extracts solutions, maintenance time, and spare parts lists, and generates a structured maintenance report that includes a 3D annotation map, maintenance priority ranking, and historical case references. The report format adopts an XML structure that conforms to marine engineering standards. The real-time terminal push unit uses the Beidou satellite communication module to establish an encrypted transmission channel for the structured maintenance report and pushes it to the maintenance personnel's terminal in real time. The push content includes a fault warning pop-up, maintenance report attachments, and a real-time communication interface. The system automatically records the push time, reception status, and maintenance personnel confirmation information, forming a closed-loop maintenance record chain.

8. A machine learning-based marine engineering data management method, applied to the machine learning-based marine engineering data management system described in any one of claims 1-7, characterized in that: Includes the following steps: S1. Deploy sensors at parts of the equipment prone to stress concentration, synchronously collect submicron vibration data and process it through spatiotemporal consistency to generate a multidimensional dataset containing time, space and amplitude dimensions. S2. Extract marine environmental noise data from the historical fault case database, reconstruct the phase space and calculate the topological features, and cluster to generate a noise feature database to provide a comparison benchmark for real-time noise separation. S3. Map the real-time vibration data to the phase space to construct a topology network, separate the noise by comparing it with the noise feature library, and then extract the fault feature vector containing geometric and dynamic characteristics based on the equipment structure model. S4. Construct a rule chain based on historical failure cases, input the extracted feature vectors and real-time environmental parameters into the rule chain for deduction, and deduce the failure type, location and severity through multi-layer logical matching. S5. Then, perform multi-source fusion verification on the inference results, calculate the rule matching degree, historical similarity and environmental robustness confidence, and trigger secondary inference to optimize the judgment result when it is lower than the threshold. S6. Generate a structured report containing 3D annotations, maintenance priorities, and historical cases based on the final fault determination. The report is then encrypted and pushed to the maintenance terminal via BeiDou communication to form a closed-loop maintenance record.

Images