Ship state monitoring system and method based on multi-source data

Abstract

The invention discloses a ship state monitoring system and method based on multi-source data, and relates to the technical field of ship engineering, the system comprises a bionic flexible sensing module, a multi-source data acquisition module, an anti-interference communication module, a multi-source data fusion module, an edge calculation processing module, an intelligent fault diagnosis module and a man-machine interaction and decision support module; wherein the bionic flexible sensing module acquires broadband vibration signals, the multi-source data acquisition module accesses multi-source data, the anti-interference communication module realizes data transmission based on a metamaterial antenna and the like, the multi-source data fusion module generates a standardized data set, and the edge calculation processing module completes feature extraction and abnormity pre-screening. The intelligent fault diagnosis module locates a fault and generates a report, and the man-machine interaction and decision support module visualizes data; the bottleneck of curved surface detection is broken through, data transmission is guaranteed, and accurate diagnosis is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ship engineering, and in particular to a ship status monitoring system and method based on multi-source data. Background Art

[0002] Existing ship condition monitoring systems generally use rigid sensors (such as piezoelectric ceramics and strain gauges) and centralized wiring architectures, which present core flaws. Rigid sensors struggle to conform to the irregular curves of ship bulkheads, such as pipe bends with a curvature radius of less than 0.5m. This results in blind spots for detecting low-frequency vibrations (0.01-0.1Hz) (such as hull deformation caused by waves) and high-frequency mechanical vibrations (500-1kHz) (such as main engine gearbox meshing noise). Measurements on a container ship show that traditional sensors experience over 30% signal attenuation in curved areas, and installation stress shortens their lifespan by 40%. In centralized wiring networks, single-point failures such as cable breaks and node failures can easily cause regional monitoring disruptions. For example, a bulk carrier experienced missed detection of early-stage gearbox cracks due to communication interruptions between engine room sensor nodes. Furthermore, traditional sensor networks based on time-of-day positioning (TDoA) have positioning errors exceeding 20cm in densely packed engine room piping, making them unable to accurately locate equipment-level fault sources such as single bearing failures.

[0003] The metal shielding and dense piping in a ship's engine room pose a significant challenge to wireless communications. Traditional technologies like Wi-Fi and LoRa have an effective transmission range of only 30-50 meters in metal cabins, and the data transmission success rate decreases significantly with distance, dropping from 80% to 50% at 50 meters. Equipment vibration and personnel movement during navigation cause frequent changes in network topology. Existing self-organizing networks, such as the AODV protocol, have routing switching delays exceeding 200ms, making them unable to meet real-time monitoring requirements. Furthermore, the complex structure of engine room piping causes multipath reflections, resulting in bit error rates exceeding 5% for traditional communication technologies, which are more pronounced in high-frequency bands exceeding 500MHz. Summary of the Invention

[0004] The object of the present invention is to provide a ship status monitoring system and method based on multi-source data to solve the problems raised in the prior art.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A ship state monitoring system based on multi-source data includes a bionic flexible sensing module, a multi-source data acquisition module, an anti-interference communication module, a multi-source data fusion module, an edge computing processing module, an intelligent fault diagnosis module, and a man-machine interaction and decision support module. The bionic flexible sensing module is distributedly integrated in the ship bulkhead for collecting ship bulkhead broadband vibration signals, covering low to high frequency vibration bands. The vibration signals are synchronously accessed to the system through the multi-source data acquisition module. The multi-source data acquisition module is used for synchronously accessing ship sensors, external data sources, and bionic flexible sensing module data. The anti-interference communication module is based on metamaterial antennas and frequency hopping spread spectrum technology, combined with an ant colony optimization dynamic routing algorithm, for data transmission in a metal shielding and pipeline dense environment. The multi-source data fusion module aligns different sampling rate data through a time synchronization algorithm, unifies heterogeneous data formats using protocol-independent analysis technology, and generates a standardized monitoring data set. The edge computing processing module deploys a lightweight AI model on the engine room edge node to process vibration signals and equipment operating parameter data in real time, complete feature extraction and abnormal pre-screening. The intelligent fault diagnosis module constructs a device health model based on multi-modal data, combines a knowledge graph and an interpretable AI algorithm, locates multi-component coupled faults, and generates a diagnosis report and maintenance recommendations. The man-machine interaction and decision support module visualizes ship state data through an augmented reality interface, allowing crew members to view equipment operating parameters, historical case comparisons, and emergency response plans.

[0007] The bionic flexible module includes a sensing unit and a signal optimization unit.

[0008] The sensing unit is responsible for real-time collection of ship bulkhead broadband vibration signals, covering the full frequency band from low to high frequency. Specifically, a flexible sensing element is designed to convert bulkhead vibration into electrical signals through piezoelectric effect, and the signal collection range covers low-frequency structural vibration and high-frequency mechanical vibration generated during ship operation. Then, an array layout is used to expand the monitoring coverage, enabling real-time acquisition of full-ship vibration data. The collected signals are synchronously accessed to the system through the interface of the multi-source data acquisition module.

[0009] The signal optimization unit is used to improve the strain sensitivity of low-frequency vibration signals, suppress high-frequency structural noise, and improve the separation degree of signals and noise, providing high signal-to-noise ratio vibration data for subsequent data processing.

[0010] Specifically, for low-frequency enhancement: by optimizing the stress conduction path, the strain energy generated by low-frequency vibration is more concentratedly transmitted to the sensing element, improving the electrical signal output intensity under unit vibration amplitude, thereby enhancing the sensitivity of low-frequency signals.

[0011] High-frequency noise suppression: by introducing a frequency-selective damping mechanism, the stress generated by high-frequency environmental noise is scattered, reducing its interference with the sensing element and lowering the noise level of high-frequency signals.

[0012] The optimized signal is processed by a protocol analysis unit of the multi-source data acquisition module, converted into a standardized data stream in a unified format, and provided for time synchronization and noise filtering by the multi-source data fusion module.

[0013] The multi-source data acquisition module comprises a data access unit and a protocol analysis unit.

[0014] The data access unit is provided with a CAN bus, a Modbus RTU, a NMEA 0183 serial port and an Ethernet interface, which are respectively connected to the sensors of the engine room equipment, the industrial controller, the navigation equipment and the external data source, and realize electrical characteristic matching through hardware circuit to obtain the state data of the corresponding equipment in real time.

[0015] The protocol analysis unit is internally provided with a CANopen protocol analysis program, a Modbus RTU protocol analysis program, a NMEA 0183 protocol analysis program and an OPC UA protocol analysis program. The CANopen protocol analysis program analyzes the device identifier and object dictionary data, the Modbus RTU protocol analysis program processes the register address and function code information, the NMEA 0183 protocol analysis program extracts the statement header and field data, and the OPC UA protocol analysis program analyzes the information model and service request. Each analysis program identifies the original byte stream transmitted by the data access unit frame by frame through the preset protocol rule, extracts the device number, parameter name and data value information, and generates a standardized data stream in a unified JSON format for calling by the multi-source data fusion module.

[0016] The anti-interference communication module comprises a metamaterial antenna unit, a frequency hopping spread spectrum processing unit and an ant colony optimization routing unit.

[0017] The metamaterial antenna unit enhances the ability of electromagnetic waves to penetrate metal shielding environments through resonance frequency regulation technology of artificial electromagnetic structures, realizes efficient transmission of signals in complex metal cabins, and its resonance frequency is determined through electromagnetic simulation optimization to adapt to the strong interference environment of the ship engine room.

[0018] The frequency hopping spread spectrum processing unit integrates a pseudo-random sequence generator, which generates frequency control instructions according to a preset frequency hopping pattern to control the communication module to switch the working frequency band at fixed time intervals, and simultaneously performs spread spectrum modulation on the transmission data by multiplying the to-be-transmitted signal with the pseudo-random sequence to expand it to a wide frequency band. The receiving end performs despreading processing through the same pseudo-random sequence to complete data recovery.

[0019] The ant colony optimization routing unit is provided with a monitoring module at each communication node to collect signal strength and transmission delay parameters in real time.

[0020] Specifically: first, the link quality is determined by a preset threshold, and the routing reconstruction process is triggered when the parameter exceeds the threshold; then the node prioritizes the neighbor nodes to select relay nodes based on the dynamic evaluation of pheromone concentration and path loss, wherein the path loss is evaluated by weighting the signal strength and transmission delay parameters, specifically, the signal strength threshold and transmission delay threshold are set as the reference, the difference between the real-time monitored signal strength, transmission delay and the corresponding threshold is respectively assigned a weight coefficient, the signal strength weight accounts for a higher proportion to preferentially guarantee the communication quality, and the path loss value is obtained by weighted summation, the smaller the value, the better the path quality, which provides a quantitative basis for relay node selection; the pheromone concentration is dynamically updated based on the path transmission state, when the path successfully transmits data, the pheromone is strengthened by a fixed increment to mark the high-quality path, and if the transmission fails, the pheromone concentration is reduced by a preset attenuation coefficient to weaken the poor path, through this positive feedback mechanism, the node is guided to preferentially select the node with high pheromone concentration and low path loss when selecting the relay node, thereby continuously optimizing the communication path; finally, the communication path is reconstructed by broadcasting route request frames, unicast route response frames and return confirmation frames.

[0021] The multi-source data fusion module comprises a time synchronization unit and a data standardization unit;

[0022] The time synchronization unit adopts a combination of hardware triggering and software compensation, the hardware level obtains coordinated universal time as the reference time through the GPS time service module, and each data source node is equipped with a crystal oscillator clock to realize initial time alignment through a synchronization pulse signal; the software level uses Kalman filtering algorithm to estimate and compensate the clock offset in real time, and the state equation is as follows:

[0023] X k =AX k-1 +W k ,

[0024] Wherein X k represents the state vector at time k, including clock offset and drift rate, A is the state transition matrix, and W k is the process noise; the measurement equation is as follows:

[0025] Z k =HX k +V k ,

[0026] Wherein Z k is the measurement value at time k, H is the observation matrix, and V k is the measurement noise, through state prediction and update iteration calculation, the time deviation of each data source is minimized;

[0027] The data standardization unit identifies the format and semantics of the original data through protocol-independent parsing technology, uses a finite state machine to parse the data frame byte by byte, extracts the protocol identification field, and then matches the corresponding parsing rules to map the data of different protocols to the pre-defined data model; then the unit uniform conversion is carried out on the parsed data, the unit conversion table is established for physical quantity, and the unit normalization is realized by table lookup method; the sliding window filtering algorithm is used for smoothing the data, and the filtering formula of window size n is as follows:

[0028] Y i =(X i-m +X i-m+1 +…+X i +…+X i+m-1 +X i+m ) / n,

[0029] Where X i is the original data sequence, Y i is the filtered data sequence, m is the window radius, and the random noise interference is eliminated by weighted average, and finally the standardized monitoring data set containing device number, parameter name, time stamp and value is generated, and the field structure conforms to the pre-defined data model.

[0030] The edge computing processing module includes a model deployment unit and a data processing unit.

[0031] The model deployment unit selects a lightweight AI model architecture MobileNet according to the calculation resource limitation of the ship engine room edge node, removes redundant convolution layers through structure pruning, quantizes the model parameters to compress the precision from 32-bit floating point to 8-bit integer, then combines the knowledge distillation technology to migrate the knowledge of complex models to lightweight models; then using TensorFlow Lite inference framework, the compressed model file is converted into an executable format for edge nodes, the instruction set architecture of embedded processors is adapted through cross-compilation tool chain, the mapping relationship between model input tensor and standardized data field is established, and the environment parameters required for model running are configured, completing the initialization and deployment of the model in the edge node;

[0032] The data processing unit monitors the standardized data stream output by the multi-source data fusion module in real time, calculates the time-frequency distribution matrix of the signal through a short-time Fourier transform algorithm for the vibration signal data field, selects a Hanning window as the window function of the short-time Fourier transform, sets the window length as an integer multiple of the signal period, generates a three-dimensional feature matrix containing time-frequency-energy density, and inputs the three-dimensional feature matrix into a deployed convolutional neural network (CNN) model; the CNN model adopts a cascade structure of a 3x3 convolution kernel and a 2x2 pooling kernel, layer by layer extracts the frequency domain texture features, modulation features and fault sensitive features of the vibration signal, and finally outputs a feature vector with a dimension of 1xN; for the equipment operation parameters, a sliding window algorithm with a fixed window size is adopted to calculate the mean, variance, kurtosis and peak factor statistical features of the data in the window at a second level, and then an M-dimensional feature vector is generated and input into a lightweight support vector machine (SVM) model; the SVM model adopts a radial basis kernel function, and the kernel function formula is as follows:

[0033] K(xi,xj)=exp(-γ||xi-xj||2),

[0034] where xi and xj represent two sample vectors in the feature space, and γ is a kernel width parameter determined through cross-validation.

[0035] Then the feature vector is mapped to a high-dimensional space, and a classification hyperplane is solved by using a soft margin maximization principle, and the decision function is as follows:

[0036] f(x)=sign(w·φ(x)+b),

[0037] where w is the weight vector of the optimal classification hyperplane, φ(x) is the feature space mapping function, and b is the bias term. The abnormal state is pre-screened by calculating the distance of the sample to the hyperplane, and finally the vibration feature vector with a dimension of 1xN output by the CNN and the M-dimensional parameter feature pre-screening result generated by the SVM are packaged into a JSON format intermediate data frame containing the feature vector and a screening state field, and are transmitted to the intelligent fault diagnosis module through an Ethernet interface.

[0038] The intelligent fault diagnosis module includes a knowledge graph construction unit, a health degree modeling unit, an interpretable reasoning unit and a diagnosis report generation unit.

[0039] The knowledge graph construction unit acquires ship equipment manuals and historical fault cases, identifies equipment components, fault types and feature parameters by using entity extraction technology, constructs causal relationships between entities by using a relationship extraction algorithm, forms a knowledge graph containing nodes and directed edges, wherein the nodes represent entities and the directed edges represent the influence relationship between entities, and adopts a graph database to store and manage the knowledge graph data.

[0040] The health degree modeling unit receives the vibration signal feature vector output by the edge computing processing module and the equipment operation parameter feature vector, wherein the vibration signal feature vector is output by the convolutional neural network and has a dimension of 1xN and contains frequency domain features extracted from the vibration signal, and the equipment operation parameter feature vector is generated by the support vector machine preprocessing and has a dimension of M and contains statistical features of the operation parameters; a fusion neural network model is used to process the two types of feature vectors, which are transmitted to the full connection layer as input, and the feature fusion is realized through the weight matrix and bias vector of the full connection layer, and the calculation formula of the specific fusion model is as follows:

[0041]

[0042] wherein F v is the vibration feature vector, F p is the parameter feature vector, W1 and W2 are weight matrices corresponding to the vibration feature vector and the parameter feature vector respectively, used to adjust the influence degree of different features, b is a bias vector used to translate and adjust the linear combination result, is an activation function, which performs nonlinear transformation on the linear combination result, compresses the value range to (0, 1), and finally converts the activation function output result to a device health degree score S of (0-100) through linear scaling, which comprehensively reflects the device running state, and the numerical value reflects the good or bad of the device health state;

[0043] The interpretable reasoning unit inputs the abnormal state with a health degree score lower than a preset threshold into the knowledge graph, matches the fault rule chain corresponding to the abnormal features through the breadth-first search algorithm, calculates the posterior probability of each fault node in combination with the Bayesian network, and generates a reasoning process containing a fault path, wherein the conditional probability table of the Bayesian network is obtained based on historical fault data statistics;

[0044] The diagnostic report generation unit automatically extracts the fault component name, fault type and characteristic parameter deviation value information according to the reasoning result, calls a predefined report template to generate a structured document, and retrieves a maintenance scheme of a similar historical case from the knowledge graph as a maintenance suggestion for the current fault, and the report content contains fault positioning results, reasoning basis and maintenance steps, which are transmitted to the man-machine interaction and decision support module through an Ethernet interface.

[0045] The man-machine interaction and decision support module includes a data visualization unit, an interactive control unit, a historical case management unit and an emergency disposal support unit.

[0046] The data visualization unit builds a ship digital twin model corresponding to the actual structure and equipment location of the ship based on augmented reality technology through the Unity 3D engine; then the equipment health score, vibration feature vector and operating parameters output by the intelligent fault diagnosis module are associated to the corresponding components in the digital twin model according to the preset mapping relationship; then the abnormal area is visualized in the form of a heat map, and the abnormal degree of the equipment state is represented by a color gradient; finally, the digital twin model and the data visualization results are rendered in real time on the browser end using WebGL technology, and the crew can view the distribution of the equipment state of the whole ship through the terminal interface and obtain intuitive ship state monitoring information.

[0047] The interactive control unit integrates a touch input and a speech recognition module, wherein the touch input adopts a capacitive touch screen controller supporting multi-point touch operation; then according to a gesture recognition algorithm, sliding and zooming gestures are realized to control the monitoring interface and to retrieve detailed data of a single device; a speech recognition module based on a hidden Markov model is constructed, a special speech instruction set for the ship field is pre-trained, speech signals are collected through a microphone array and converted into control instructions; at the same time, according to an instruction analysis engine, the speech instructions are mapped to corresponding system operations; finally, the real-time transmission of interactive instructions and system responses are realized through the WebSocket protocol.

[0048] The historical case management unit connects a MySQL database to store structured historical fault data, performs keyword search through a Lucene search engine, and supports multi-dimensional filtering according to equipment types and fault types; the staff can compare the similarity of the current abnormal state with the characteristics of the historical cases to assist in judging the fault mode;

[0049] The emergency disposal support unit has a pre-defined report template library built-in, automatically extracts fault information in the diagnostic report to fill in the template to generate a document, searches the maintenance knowledge base in the knowledge graph, displays the standard maintenance process corresponding to the fault, and guides the maintenance operation.

[0050] A ship state monitoring method based on multi-source data, comprising the following steps:

[0051] S1, the system collects ship bulkhead broadband vibration signals through a bionic flexible sensing module, and simultaneously uses a multi-source data collection module to synchronously access data from ship sensors and external data sources;

[0052] S2, the collected data is processed by an anti-interference communication module, based on a metamaterial antenna, frequency hopping spread spectrum technology and an ant colony optimization dynamic routing algorithm, to complete transmission in a metal shielding and pipeline dense environment;

[0053] S3, the multi-source data after transmission enters the multi-source data fusion module, the different sampling rate data are aligned through a time synchronization algorithm, and the data format is unified through a protocol-independent analysis technology, and a standardized monitoring data set is generated;

[0054] S4, the standardized data set is transmitted to an edge computing processing module, the vibration signal and the device operation parameter are extracted and abnormally pre-screened through a lightweight AI model in the engine room edge node;

[0055] S5, the result after edge computing processing is input into an intelligent fault diagnosis module, a device health degree model is constructed based on multi-modal data, a knowledge graph and an interpretable AI algorithm are combined, a fault is located, and a diagnosis report and a maintenance suggestion are generated;

[0056] S6, the result output by the intelligent fault diagnosis module is presented through a man-machine interaction and a decision support module, the ship state data is visualized in an augmented reality interface, the parameters are viewed by the staff, the cases are compared, and the emergency disposal plan is obtained.

[0057] Compared with the prior art, the beneficial effects of the present application are:

[0058] 1, the flexible sensing technology breaks through the bottleneck of curved surface detection: the flexible characteristics of the PVDF piezoelectric film can adapt to the installation of the complex curved surface of the ship, eliminate the installation stress of the traditional rigid sensor, and prolong the service life of the sensor; the fishbone structure concentrates the stress to the center through the radial grooves, improves the strain sensitivity in the low frequency band (0.01-0.1 Hz), suppresses the structure noise in the high frequency band (500-1 kHz), realizes the non-blind area collection of the wide frequency vibration signal (covering the full frequency band from low frequency to high frequency), and reduces the signal attenuation rate.

[0059] 2, anti-interference communication guarantees data transmission: the multi-layer composite structure of the metal split ring resonator and the dielectric substrate enhances the ability of the specific frequency electromagnetic wave to penetrate the metal shield, the effective transmission distance is extended to 80-100 meters, and the signal penetration loss is reduced by more than 20 dB; the frequency hopping spread spectrum technology switches the frequency band through a pseudo-random sequence and spread spectrum modulation, reduces the bit error rate caused by multi-path reflection to below 1%; the ant colony optimization routing unit dynamically updates the information through the pheromone mechanism, real-time reconstructs the communication link, shortens the routing switching delay, and ensures the continuous transmission of data in the engine room environment with frequent topology changes.

[0060] 3. AI analysis realizes accurate diagnosis: the ship digital twin model combines the heat map to display the equipment health degree in real time, the abnormal area is intuitively presented through the color gradient, the crew can quickly locate the equipment state of the whole ship, the decision-making efficiency is improved; voice recognition and gesture control support touchless operation, adapt to the humid and oily environment of the engine room; the historical case management unit provides similar fault maintenance scheme comparison through keyword retrieval and multi-dimensional screening, and the emergency disposal time is shortened to 30% of the traditional process. BRIEF DESCRIPTION OF DRAWINGS

[0061] Fig. 1 The figure is a organizational architecture diagram of a ship state monitoring system based on multi-source data according to the present application;

[0062] Fig. 2 The figure is a work flow diagram of a ship state monitoring system based on multi-source data according to the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0064] Embodiment: as shown in the figure, the present application provides a technical solution, Figs. 1-2

[0065] ​A ship state monitoring system based on multi-source data includes a bionic flexible sensing module, a multi-source data acquisition module, an anti-interference communication module, a multi-source data fusion module, an edge computing processing module, an intelligent fault diagnosis module, and a man-machine interaction and decision support module. The bionic flexible sensing module is distributedly integrated in the ship bulkhead for collecting ship bulkhead broadband vibration signals, covering low to high frequency vibration frequency bands. The vibration signals are synchronously accessed to the system through the multi-source data acquisition module. The multi-source data acquisition module is used for synchronously accessing ship sensors, external data sources, and bionic flexible sensing module data. The anti-interference communication module is based on metamaterial antenna and frequency hopping spread spectrum technology, combined with ant colony optimization dynamic routing algorithm, for data transmission in a metal shielding and pipeline dense environment. The multi-source data fusion module aligns different sampling rate data through a time synchronization algorithm, unifies heterogeneous data formats using protocol-independent analysis technology, and generates a standardized monitoring data set. The edge computing processing module deploys a lightweight AI model on the engine room edge node to process vibration signals and equipment operating parameter data in real time, complete feature extraction and abnormal pre-screening. The intelligent fault diagnosis module constructs a device health degree model based on multi-modal data, combines knowledge graph and explainable AI algorithms, locates multi-component coupled faults, and generates diagnosis reports and maintenance recommendations. The man-machine interaction and decision support module visualizes ship state data through an augmented reality interface, allowing crew members to view equipment operating parameters, historical case comparisons, and emergency response plans.

[0066] The bionic flexible module includes a sensing unit and a signal optimization unit.

[0067] The sensing unit is responsible for real-time collection of ship bulkhead broadband vibration signals, covering the full frequency band from low to high frequency. Specifically, a flexible sensing element is designed to convert bulkhead vibration into electrical signals through piezoelectric effect, and the signal collection range covers low-frequency structural vibration and high-frequency mechanical vibration generated during ship operation. Then, an array layout is used to expand the monitoring coverage, realizing real-time acquisition of full-ship vibration data. The collected signals are synchronously accessed to the system through the interface of the multi-source data acquisition module.

[0068] The signal optimization unit is used to improve the strain sensitivity of low-frequency vibration signals, suppress high-frequency structural noise, and improve the separation degree of signals and noise, providing high signal-to-noise ratio vibration data for subsequent data processing.

[0069] Specifically, for low-frequency enhancement: by optimizing the stress conduction path, the strain energy generated by low-frequency vibration is more concentratedly transmitted to the sensing element, improving the electrical signal output intensity under unit vibration amplitude, thereby enhancing the sensitivity of low-frequency signals.

[0070] High-frequency noise suppression: by introducing a frequency-selective damping mechanism, the stress generated by high-frequency environmental noise is scattered, reducing its interference with the sensing element and lowering the noise level of high-frequency signals.

[0071] The optimized signal is processed by the protocol analysis unit of the multi-source data acquisition module, converted into a standardized data stream in a unified format, and provided to the multi-source data fusion module for time synchronization and noise filtering;

[0072] The unit compensates for the detection blind area of traditional sensors through wideband signal acquisition and frequency band optimization, enhances the ability to capture early fault features such as low-frequency deformation and high-frequency impact, provides multi-dimensional feature input for the AI model of the edge computing module, and supports precise positioning of the intelligent fault diagnosis module.

[0073] The multi-source data acquisition module includes a data access unit and a protocol analysis unit;

[0074] The data access unit is provided with a CAN bus, a Modbus RTU, a NMEA 0183 serial port, and an Ethernet interface, which are respectively connected to the engine room equipment sensors, industrial controllers, navigation equipment, and external data sources. The electrical characteristics are matched through hardware circuit to obtain the state data of the corresponding equipment in real time;

[0075] The protocol analysis unit is built-in CANopen protocol analysis program, Modbus RTU protocol analysis program, NMEA 0183 protocol analysis program and OPC UA protocol analysis program. The CANopen protocol analysis program analyzes the device identifier and object dictionary data, the Modbus RTU protocol analysis program processes the register address and function code information, the NMEA 0183 protocol analysis program extracts the statement header and field data, and the OPC UA protocol analysis program analyzes the information model and service request. Each analysis program identifies the original byte stream transmitted by the data access unit frame by frame through the preset protocol rule, extracts device number, parameter name, data value and other information, generates a standardized data stream in a unified JSON format, and calls the multi-source data fusion module.

[0076] The anti-interference communication module includes a metamaterial antenna unit, a frequency hopping spread spectrum processing unit, and an ant colony optimization routing unit;

[0077] The metamaterial antenna unit enhances the ability of electromagnetic waves to penetrate metal shielding environments through resonance frequency regulation technology of artificial electromagnetic structures, achieving efficient transmission of signals in complex metal cabins. The resonance frequency is determined through electromagnetic simulation optimization to adapt to the strong interference environment of the ship engine room.

[0078] The frequency hopping spread spectrum processing unit integrates a pseudo-random sequence generator, which generates frequency control instructions according to a preset frequency hopping pattern, controls the communication module to switch the working frequency band at fixed time intervals, and simultaneously performs spread spectrum modulation on the transmission data, so as to multiply and expand the to-be-transmitted signal to a wide frequency band by the pseudo-random sequence; the receiving end performs despreading processing by using the same pseudo-random sequence to complete data recovery;

[0079] The ant colony optimization routing unit is provided with a monitoring module at each communication node to collect signal strength and transmission time delay parameters in real time.

[0080] Specifically, first, the link quality is determined by a preset threshold, and when the parameters exceed the threshold, the routing reconstruction process is triggered; then, the node performs priority sorting on the neighbor nodes based on the dynamic evaluation of pheromone concentration and path loss to select a relay node, wherein the path loss is evaluated by weighting the signal strength and transmission time delay parameters, specifically, the signal strength threshold and the transmission time delay threshold are set as the reference, the difference between the real-time monitored signal strength and transmission time delay and the corresponding threshold is respectively assigned a weight coefficient, the signal strength weight accounts for a higher proportion to preferentially guarantee the communication quality, and then the path loss value is obtained by weighted summation, and the smaller the value, the better the path quality, which provides a quantitative basis for relay node selection; the pheromone concentration is dynamically updated based on the path transmission state, when the path successfully transmits data, the pheromone is strengthened by a fixed increment to mark the high-quality path, and if the transmission fails, the pheromone concentration is reduced by a preset attenuation coefficient to weaken the poor path, through this positive feedback mechanism, the node is guided to preferentially select the node with high pheromone concentration and low path loss when selecting a relay node, thereby continuously optimizing the communication path; finally, the communication path is reconstructed by broadcasting the route request frame, the unicast route response frame and the return confirmation frame;

[0081] The unit ensures that the vibration data of the bionic flexible sensing module and the sensor parameters of the multi-source data acquisition module are stably transmitted to the multi-source data fusion module through the anti-interference communication module, avoids communication interruption caused by engine compartment equipment vibration, provides continuous data source for edge computing and fault diagnosis, and guarantees the real-time performance and reliability of the entire monitoring system.

[0082] The multi-source data fusion module comprises a time synchronization unit and a data standardization unit.

[0083] The time synchronization unit adopts a combination of hardware triggering and software compensation, the hardware level obtains coordinated universal time as the reference time through the GPS time service module, and each data source node is provided with a crystal oscillator clock to realize initial time alignment through a synchronization pulse signal; the software level uses the Kalman filtering algorithm to estimate and compensate the clock offset in real time, and the state equation is as follows:

[0084] X k =AX k-1+W k ,

[0085] where X k represents the state vector at time k, containing clock offset and drift rate, A is the state transition matrix, W k is the process noise; the measurement equation is as follows:

[0086] Z k = HX k + V k ,

[0087] where Z k is the measurement value at time k, H is the observation matrix, V k is the measurement noise, through state prediction and update iteration calculation, the minimization of each data source time deviation is realized;

[0088] The data standardization unit identifies the format and semantics of the original data through protocol-independent analysis technology, uses a finite state machine to parse the data frame byte by byte, extracts the protocol identification field and matches the corresponding analysis rules, maps the data of different protocols to the pre-defined data model; then the unit conversion of the parsed data is carried out, the unit conversion table is established for physical quantities, and the unit normalization is realized by table lookup method; the sliding window filtering algorithm is used for smoothing the data, and the filtering formula with window size n is as follows:

[0089] Y i = (X i-m + X i-m+1 + … + X i + … + X i+m-1 + X i+m ) / n,

[0090] where X i is the original data sequence, Y i is the filtered data sequence, m is the window radius, the random noise interference is eliminated by weighted average, and finally the standardized monitoring data set containing device number, parameter name, time stamp and value is generated, and the field structure conforms to the pre-defined data model.

[0091] The edge computing processing module includes a model deployment unit and a data processing unit;

[0092] The model deployment unit selects a lightweight AI model architecture MobileNet for the computing resource limit of the ship engine room edge node, removes redundant convolution layers through structural pruning, quantizes the model parameters to compress the precision from 32-bit floating point to 8-bit integer, then migrates the knowledge of the complex model to the lightweight model combined with the knowledge distillation technology; then the compressed model file is converted into an executable format for the edge node using the TensorFlow Lite inference framework, the instruction set architecture of the embedded processor is adapted through the cross-compilation tool chain, the mapping relationship between the model input tensor and the standardized data field is established, and the environment parameters required for model running are configured to complete the initialization and deployment of the model on the edge node.

[0093] The data processing unit real-time monitors the standardized data stream output by the multi-source data fusion module, and for the vibration signal data field, calculates the time-frequency distribution matrix of the signal through the short-time Fourier transform algorithm, wherein the window function of the short-time Fourier transform is selected as the Hanning window, and the window length is set as an integer multiple of the signal period, a three-dimensional feature matrix containing time-frequency-energy density is generated, and input to the deployed convolutional neural network model CNN; the CNN model adopts a cascade structure of 3x3 convolution kernel and 2x2 pooling kernel, and extracts the frequency domain texture features, modulation features and fault sensitive features of the vibration signal layer by layer, and finally outputs a feature vector with a dimension of 1xN; for the equipment operating parameters, a sliding window algorithm with a fixed window size is used to calculate the mean, variance, kurtosis and peak factor statistical features of the data in the window at a second level, and then input the M-dimensional feature vector to the lightweight support vector machine model SVM; the SVM model adopts a radial basis kernel function, and the kernel function formula is as follows:

[0094] K(xi,xj)=exp(-γ||xi-xj||2),

[0095] wherein xi and xj represent two sample vectors in the feature space, and γ is a kernel width parameter determined through cross-validation;

[0096] Then the feature vector is mapped to a high-dimensional space, and the soft margin maximization principle is used to solve the classification hyperplane, and the decision function is as follows:

[0097] f(x)=sign(w·φ(x)+b),

[0098] wherein w is the weight vector of the optimal classification hyperplane, φ(x) is the feature space mapping function, and b is the bias term, and the distance from the sample to the hyperplane is calculated to realize the pre-screening of the abnormal state, and finally the vibration feature vector with a dimension of 1xN output by the CNN and the M-dimensional parameter feature pre-screening result generated by the SVM are packaged into a JSON format intermediate data frame containing the feature vector and the screening state field, and transmitted to the intelligent fault diagnosis module through the Ethernet interface.

[0099] The intelligent fault diagnosis module comprises a knowledge graph construction unit, a health degree modeling unit, an interpretable reasoning unit and a diagnosis report generation unit.

[0100] The knowledge graph construction unit identifies equipment components, fault types and characteristic parameters by using entity extraction technology through collecting ship equipment manuals and historical fault cases, constructs causal relationships between entities by using a relationship extraction algorithm, forms a knowledge graph containing nodes and directed edges, wherein the nodes represent entities and the directed edges represent the influence relationship between entities, and adopts a graph database to store and manage knowledge graph data.

[0101] The health degree modeling unit receives the vibration signal feature vector and the equipment operating parameter feature vector output by the edge computing processing module, wherein the vibration signal feature vector is output by a convolutional neural network and has a dimension of 1xN, containing frequency domain features extracted from the vibration signal, and the equipment operating parameter feature vector is generated by a support vector machine preprocessing and has a dimension of M, containing statistical features of operating parameters; a fusion neural network model is used to process the two types of feature vectors, which are transmitted to the full connection layer as input, and the feature fusion is realized through the weight matrix and bias vector of the full connection layer. The calculation formula of the specific fusion model is as follows:

[0102]

[0103] wherein F v is the vibration feature vector, F p is the parameter feature vector, W1 and W2 are weight matrices corresponding to the vibration feature vector and the parameter feature vector respectively, used to adjust the influence degree of different features, b is a bias vector used to translate and adjust the linear combination result, is an activation function, which performs a nonlinear transformation on the linear combination result to compress the value range to (0, 1), and finally converts the activation function output result to a device health degree score S of (0-100) through linear scaling, which comprehensively reflects the equipment operating state, and the numerical value reflects the good or bad of the equipment health state;

[0104] The interpretable reasoning unit inputs the abnormal state with a health degree score lower than a preset threshold into the knowledge graph, matches the fault rule chain corresponding to the abnormal feature by using a breadth-first search algorithm, calculates the posterior probability of each fault node by combining a Bayesian network, and generates a reasoning process containing a fault path, wherein the conditional probability table of the Bayesian network is obtained based on historical fault data statistics;

[0105] The diagnostic report generation unit automatically extracts the fault component name, fault type, and characteristic parameter deviation value information based on the reasoning results, calls a predefined report template to generate a structured document, and simultaneously retrieves maintenance plans for similar historical cases from the knowledge graph as maintenance recommendations for the current fault. The report content includes the fault location results, reasoning basis, and maintenance steps, and is transmitted to the human-computer interaction and decision support module via the Ethernet interface.

[0106] The human-computer interaction and decision support module includes a data visualization unit, an interaction control unit, a historical case management unit and an emergency response support unit;

[0107] The data visualization unit uses augmented reality technology to construct a digital twin model of the ship using the Unity 3D engine. This model corresponds one-to-one with the actual ship structure and equipment locations. The equipment health score, vibration eigenvector, and operating parameters output by the intelligent fault diagnosis module are then associated with the corresponding components in the digital twin model according to a preset mapping relationship. Abnormal areas are then visualized in the form of a heat map, with color gradients used to represent the degree of abnormality in the equipment status. Finally, WebGL technology is used to enable real-time browser-side rendering of the digital twin model and data visualization results. Crew members can view the status distribution of all equipment on the ship through the terminal interface, obtaining intuitive ship status monitoring information.

[0108] The interactive control unit integrates touch input and voice recognition modules, where the touch input uses a capacitive touch screen controller that supports multi-touch operation. Then, based on the gesture recognition algorithm, it realizes the control of the monitoring interface through sliding and zooming gestures, as well as the retrieval of detailed data of a single device. Then, it constructs a voice recognition module based on the hidden Markov model, pre-trains a voice command set dedicated to the marine field, collects voice signals through a microphone array and converts them into control commands. At the same time, based on the command parsing engine, the voice commands are mapped to corresponding system operations. Finally, the real-time transmission of interactive commands and system response are achieved through the WebSocket protocol.

[0109] The historical case management unit is connected to the MySQL database to store structured historical fault data. It can perform keyword retrieval through the Lucene search engine and support multi-dimensional screening by device type and fault type. Staff can compare the feature similarity of the current abnormal state with historical cases to assist in determining the fault mode.

[0110] The emergency response support unit has a built-in predefined report template library, which automatically extracts fault information from the diagnostic report and fills it into the template to generate a document. At the same time, it retrieves the maintenance knowledge base in the knowledge graph, displays the standard maintenance process corresponding to the fault, and guides maintenance operations.

[0111] A ship state monitoring method based on multi-source data, comprising the following steps:

[0112] S1, the system collects ship bulkhead broadband vibration signals through a bionic flexible sensing module, and simultaneously accesses data of ship sensors and external data sources through a multi-source data collection module;

[0113] S2, the collected data is processed by an anti-interference communication module, and is transmitted in a metal shielding and pipeline dense environment based on a metamaterial antenna, frequency hopping spread spectrum technology and an ant colony optimization dynamic routing algorithm;

[0114] S3, the multi-source data after transmission enters a multi-source data fusion module, different sampling rate data are aligned through a time synchronization algorithm, and a protocol-independent analysis technology is used to unify the data format, thereby generating a standardized monitoring data set;

[0115] S4, the standardized data set is transmitted to an edge computing processing module, data is processed in real time through a lightweight AI model at a cabin edge node, and feature extraction and abnormal pre-screening of vibration signals and equipment operation parameters are completed;

[0116] S5, the result after edge computing processing is input into an intelligent fault diagnosis module, an equipment health degree model is constructed based on multi-modal data, a knowledge graph and an interpretable AI algorithm are combined, faults are located, and a diagnosis report and maintenance suggestions are generated;

[0117] S6, the result output by the intelligent fault diagnosis module is presented through a man-machine interaction and decision support module, a ship state data is visualized in an augmented reality interface, parameters are viewed by a worker, cases are compared, and an emergency disposal plan is obtained.

[0118] Taking a container ship cabin monitoring scene as an example, the specific application process and data implementation of the system are described;

[0119] At the level of data acquisition and communication, the bionic flexible sensing module collects 0.01-1000 Hz broadband vibration signals through 50 distributed polyvinylidene fluoride (PVDF) piezoelectric film sensors. The strain sensitivity of the fishbone stress concentration structure is improved to 2 times that of traditional sensors in the low frequency band (0.01-0.1 Hz), and the noise suppression ratio reaches 15 dB in the high frequency band (500-1000 Hz). The multi-source data acquisition module connects 30 cabin equipment sensors (such as temperature and pressure sensors) through the CAN bus interface and 10 valve controllers through the Modbus RTU interface, and collects real-time parameters such as speed and oil pressure. The original data is processed by the protocol analysis unit and output in a unified JSON format, with a data transmission rate of 10 Mbps. In the anti-interference communication module, the metamaterial antenna uses a 3-layer metal open ring resonator and FR4 dielectric substrate stacking structure, with a resonant frequency of 2.4 GHz. The penetration loss of the metal cabin wall environment is reduced by 20 dB compared to traditional antennas, and the effective transmission distance reaches 80 meters. The frequency hopping spread spectrum processing unit switches the frequency band every 100 ms, with a pseudo-random sequence length of 127 bits, expanding the data to 20 MHz bandwidth transmission, and the bit error rate is controlled within 0.5%. The ant colony optimization routing unit sets the link signal strength threshold to -80 dBm, and when the link quality is lower than the threshold, the routing is rebuilt within 50 ms through the pheromone update mechanism.

[0120] In the data fusion and edge computing stage, the multi-source data fusion module obtains coordinated universal time (UTC) through the GPS timing module, and the initial synchronization error of each node crystal oscillator clock is less than 1 μs. After compensation by the Kalman filter algorithm, the time deviation is stabilized within 100 ns, where the state vector X k is 2-dimensional (offset, drift rate), the state transition matrix A = [1 1; 0 1], the process noise covariance Q = 0.01, and the measurement noise covariance R = 0.1. The data standardization unit filters the vibration signal with a 5-point sliding window (n = 5, m = 2), converts the units of temperature, pressure, and other parameters (such as psi to Pa), and generates a standardized data set containing device number, parameter value, and timestamp, with 1000 records output per second. The edge computing processing module deploys the compressed MobileNet model (parameter size 1.4 MB), which generates a 50x50 time-frequency matrix from the vibration signal through short-time Fourier transform (window length 512 ms, Hanning window), extracts features through CNN, and outputs a 1x128-dimensional feature vector. The running parameters are calculated through a sliding window to obtain 5-dimensional statistical features such as mean and variance, which are input into the SVM model (radial basis kernel function γ = 0.5) for abnormal pre-screening, with a feature vector and screening result output every 100 ms.

[0121] In the aspect of intelligent fault diagnosis and interaction, the knowledge graph construction unit collects 500 equipment manuals and 2000 historical fault cases, constructs a graph containing 1000 entities (such as "main engine gearbox" and "bearing wear") and 3000 relationships, and stores it in a Neo4j database. The health modeling unit receives the vibration feature vector and the parameter feature vector output by the edge computing, calculates the equipment health score S through the fusion neural network, wherein the weight matrix W1 (128x50) and W2 (5x50) are trained by historical data, the bias vector b = 0.1, the activation function adopts Sigmoid, and the score range is 0-100. When the health score S of a bearing is 45 (the preset threshold is 60), the explainable reasoning unit matches the fault rule chain through breadth-first search, and calculates the posterior probability of "bearing damage" as 85% by combining the Bayesian network. The fault component and the characteristic parameter deviation value (such as the vibration peak value exceeding the threshold by 30%) are automatically extracted by the diagnosis report generation unit, a repair template is called to generate a report, and 10 similar case repair steps are retrieved from the knowledge graph and pushed to the crew. The man-machine interaction module constructs a digital twin model of the engine room by Unity 3D, displays the abnormal bearing area (red represents health < 50) in a heat map, and the crew can call detailed information by touching the screen or giving voice commands "display bearing 23 vibration data". The system response time is < 1s, and the crew can complete fault disposal within 30 minutes with the help of historical case comparison and repair process visualization.

[0122] It will be obvious to a person skilled in the art that the application is not limited to the details of the above-described exemplary embodiments, but that the application can be implemented in other concrete forms without departing from the spirit or essential characteristics of the application. The embodiments should, therefore, be considered in all respects as illustrative and not restrictive, the scope of the application being defined by the appended claims rather than by the above description, and it is intended that all changes and modifications which come within the meaning and range of equivalents of the claims are encompassed thereby. 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 ship condition monitoring system based on multi-source data, characterized by: It includes a bionic flexible sensing module, a multi-source data acquisition module, an anti-interference communication module, a multi-source data fusion module, an edge computing and processing module, an intelligent fault diagnosis module and a human-computer interaction and decision support module; the bionic flexible sensing module is distributed and integrated in the ship bulkhead, and is used to collect broadband vibration signals of the ship bulkhead, covering low-frequency to high-frequency vibration frequency bands, and the vibration signals are synchronously connected to the system through the multi-source data acquisition module; the multi-source data acquisition module is used to synchronously access ship sensors, external data sources and bionic flexible sensing module data; the anti-interference communication module is based on metamaterial antennas and frequency hopping spread spectrum technology, combined with ant colony optimization dynamic routing algorithm, to transmit data in metal shielding and pipeline-dense environments; the multi-source data fusion module aligns data with different sampling rates through a time synchronization algorithm, unifies heterogeneous data formats using protocol-independent parsing technology, and generates a standardized monitoring data set; the edge computing and processing module deploys a lightweight AI model at the edge node of the cabin to process vibration signals and equipment operating parameter data in real time, and complete feature extraction and anomaly pre-screening; The intelligent fault diagnosis module builds an equipment health model based on multimodal data, combines knowledge graphs with explainable AI algorithms to locate multi-component coupling faults, and generates diagnostic reports and maintenance recommendations. The human-computer interaction and decision support module visualizes ship status data through an augmented reality interface, allowing crew members to view equipment operating parameters, historical case comparisons, and emergency response plans.

2. A ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The bionic flexible module includes a sensing unit and a signal optimization unit; The sensing unit is responsible for real-time acquisition of broadband vibration signals from the ship's bulkheads, covering the entire frequency range from low to high frequencies. Specifically, it uses a flexible sensitive element design to convert bulkhead vibrations into electrical signals through the piezoelectric effect. The signal acquisition range covers both low-frequency structural vibrations and high-frequency mechanical vibrations generated during ship operation. An array layout is then used to expand the monitoring coverage, enabling real-time acquisition of vibration data from the entire ship. The collected signals are synchronously connected to the system via the interface of the multi-source data acquisition module. The signal optimization unit is used to enhance the strain sensitivity of low-frequency vibration signals, suppress high-frequency structural noise, improve the separation between signals and noise, and provide vibration data with a high signal-to-noise ratio for subsequent data processing; Specifically: For low-frequency enhancement: By optimizing the stress conduction path, the strain energy generated by low-frequency vibration is more concentratedly transmitted to the sensing element, thereby improving the electrical signal output intensity per unit vibration amplitude, thereby enhancing the sensitivity of low-frequency signals; High-frequency noise suppression: By introducing a frequency-selective damping mechanism, the stress generated by high-frequency environmental noise is scattered, reducing its interference with the sensor element and lowering the noise level of high-frequency signals; The optimized signal is processed by the protocol parsing unit of the multi-source data acquisition module and converted into a standardized data stream in a unified format for time synchronization and noise filtering by the multi-source data fusion module.

3. The ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The multi-source data acquisition module includes a data access unit and a protocol analysis unit; The data access unit is provided with CAN bus, Modbus RTU, NMEA 0183 serial port and Ethernet interface, which are respectively connected to cabin equipment sensors, industrial controllers, navigation equipment and external data sources, and realizes electrical characteristic matching through hardware circuits to obtain status data of corresponding equipment in real time; The protocol parsing unit has built-in CANopen protocol parsing program, Modbus RTU protocol parsing program, NMEA 0183 protocol parsing program and OPC UA protocol parsing program. The CANopen protocol parsing program parses device identifiers and object dictionary data, the Modbus RTU protocol parsing program processes register addresses and function code information, the NMEA 0183 protocol parsing program extracts statement headers and field data, and the OPC UA protocol parsing program parses information models and service requests. Each parsing program uses preset protocol rules to identify the original byte stream transmitted by the data access unit frame by frame, extract information such as device number, parameter name, and data value, and generate a standardized data stream in a unified JSON format for calling by the multi-source data fusion module.

4. The ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The anti-interference communication module includes a metamaterial antenna unit, a frequency hopping spread spectrum processing unit and an ant colony optimization routing unit; The metamaterial antenna unit uses resonant frequency control technology of artificial electromagnetic structures to enhance the ability of electromagnetic waves to penetrate metal shielding environments, achieving efficient signal transmission in complex metal cabins. Its resonant frequency is determined through electromagnetic simulation optimization to adapt to the strong interference environment of the ship's engine room. The frequency hopping spread spectrum processing unit integrates a pseudo-random sequence generator, which generates frequency control instructions based on a preset frequency hopping pattern, controls the communication module to switch the operating frequency band at fixed time intervals, and simultaneously performs spread spectrum modulation on the transmitted data, multiplying the transmitted signal with the pseudo-random sequence to expand it to a wide frequency band. The receiving end performs despreading processing using the same pseudo-random sequence to complete data recovery; The ant colony optimization routing unit deploys a monitoring module at each communication node to collect signal strength and transmission delay parameters in real time; Specifically, the link quality is first determined by a preset threshold. When the parameter exceeds the threshold, the routing reconstruction process is triggered. Then, based on the dynamic evaluation of pheromone concentration and path loss, the node prioritizes neighboring nodes to select relay nodes. Path loss is weighted by comprehensively considering signal strength and transmission delay parameters. Specifically, a signal strength threshold and a transmission delay threshold are set as benchmarks, and the difference between the real-time monitored signal strength and transmission delay and the corresponding threshold is assigned a weight coefficient. A higher weight is given to signal strength to prioritize communication quality. The path loss value is then derived through weighted summation. The smaller the value, the better the path quality, providing a quantitative basis for relay node selection. The pheromone concentration is dynamically updated based on the path transmission status. When a path successfully transmits data, the pheromone concentration is increased by a fixed increment to mark the high-quality path. If the transmission fails, the pheromone concentration is reduced by a preset attenuation coefficient to weaken the low-quality path. Through this positive feedback mechanism, the node is guided to prioritize nodes with high pheromone concentration and low path loss when selecting relay nodes, thereby continuously optimizing the communication path. Finally, the communication path is reconstructed by broadcasting route request frames, unicasting route reply frames, and returning confirmation frames.

5. The ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The multi-source data fusion module includes a time synchronization unit and a data standardization unit; The time synchronization unit adopts a combination of hardware triggering and software compensation. At the hardware level, the GPS timing module obtains the coordinated universal time as the reference time. Each data source node is equipped with a crystal oscillator clock, and the initial time alignment is achieved through the synchronization pulse signal. At the software level, the Kalman filter algorithm is used to estimate and compensate the clock offset in real time. Its state equation is as follows: X k =AX k-1 +W k , where X k represents the state vector at time k, including clock offset and drift rate, A is the state transfer matrix, W k is the process noise; the measurement equation is as follows: Z k =HX k +V k , where Z k is the measurement value at time k, H is the observation matrix, V k To measure noise, the time deviation of each data source is minimized through state prediction and update iterative calculation; The data standardization unit identifies the format and semantics of the original data through protocol-independent parsing technology, uses a finite state machine to parse the data frame byte by byte, extracts the protocol identification field and matches the corresponding parsing rules, and maps the data of different protocols to a predefined data model; then, the parsed data is uniformly converted to units, a unit conversion table is established for physical quantities, and unit normalization is achieved through a table lookup method; a sliding window filtering algorithm is used to smooth the data, and the filtering formula with a window size of n is as follows: Y i =(X i-m +X i-m+1 +…+X i +…+X i+m-1 +X i+m ) / n, where X i is the original data sequence, Y i is the filtered data sequence, m is the window radius, and random noise interference is eliminated by weighted averaging. Finally, a standardized monitoring data set containing device number, parameter name, timestamp and value is generated. The field structure conforms to the predefined data model.

6. The ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The edge computing processing module includes a model deployment unit and a data processing unit; To address the computing resource constraints of edge nodes in ship engine rooms, the model deployment unit selects the lightweight AI model architecture MobileNet. It removes redundant convolutional layers through structural pruning and compresses the model parameter precision from 32-bit floating point to 8-bit integer through quantization. It then uses knowledge distillation technology to migrate the knowledge of the complex model to the lightweight model. The TensorFlow Lite inference framework is then used to convert the compressed model file into a format executable by the edge node. A cross-compilation tool chain is used to adapt to the instruction set architecture of the embedded processor, a mapping relationship between the model input tensor and the standardized data field is established, and the environmental parameters required for model operation are configured to complete the initial deployment of the model on the edge node. The data processing unit monitors the standardized data stream output by the multi-source data fusion module in real time, and calculates the time-frequency distribution matrix of the signal for the vibration signal data field using the short-time Fourier transform algorithm, wherein the window function of the short-time Fourier transform selects the Hanning window, and the window length is set to an integer multiple of the signal period, to generate a three-dimensional feature matrix containing time-frequency-energy density, which is input into the deployed convolutional neural network model CNN; the CNN model adopts a cascade structure of a 3×3 convolution kernel and a 2×2 pooling kernel to extract the frequency domain texture features, modulation features, and fault sensitivity features of the vibration signal layer by layer, and ultimately outputs a feature vector with a dimension of 1×N; for the equipment operating parameters, a sliding window algorithm with a fixed window size is used to calculate the mean, variance, kurtosis, and peak factor statistical features of the data in the window at a second-level period, generate an M-dimensional feature vector, and then input it into a lightweight support vector machine model SVM; the SVM model adopts a radial basis kernel function, and the kernel function formula is as follows: K(xi,xj)=exp(-γ||xi-xj||2), Among them, xi and xj represent two sample vectors in the feature space, and γ is the kernel width parameter, which is determined by cross-validation; Then the feature vector is mapped to a high-dimensional space, and the classification hyperplane is solved using the soft margin maximization principle. The decision function is as follows: f(x)=sign(w·φ(x)+b), Where w is the weight vector of the optimal classification hyperplane, φ(x) is the feature space mapping function, and b is the bias term. Abnormal conditions are pre-screened by calculating the distance between the sample and the hyperplane. Finally, the vibration feature vector with a dimension of 1×N output by CNN and the M-dimensional parameter feature pre-screening result generated by SVM are encapsulated into a JSON format intermediate data frame containing the feature vector and the screening status field, and transmitted to the intelligent fault diagnosis module through the Ethernet interface.

7. The ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The intelligent fault diagnosis module includes a knowledge graph construction unit, a health modeling unit, an explainable reasoning unit and a diagnosis report generation unit; The knowledge graph construction unit collects ship equipment manuals and historical fault cases, uses entity extraction technology to identify entities such as equipment components, fault types, and characteristic parameters, and uses relationship extraction algorithms to construct causal relationships between entities to form a knowledge graph containing nodes and directed edges, where nodes represent entities and directed edges represent the influence relationship between entities. A graph database is used to store and manage knowledge graph data; The health modeling unit receives the vibration signal feature vector and the equipment operation parameter feature vector output by the edge computing processing module. The vibration signal feature vector is output by the convolutional neural network, with a dimension of 1×N, and contains the frequency domain features extracted from the vibration signal. The equipment operation parameter feature vector is generated by support vector machine preprocessing, with a dimension of M, and contains the statistical characteristics of the operation parameters. The two types of feature vectors are processed using a fusion neural network model and passed as input to the fully connected layer. Feature fusion is achieved through the weight matrix and bias vector of the fully connected layer. The specific calculation formula of the fusion model is as follows: Among them F v is the vibration eigenvector, F p is the parameter eigenvector, W1 and W2 are the weight matrices corresponding to the vibration eigenvector and parameter eigenvector, respectively, which are used to adjust the influence of different features. b is the bias vector, which is used to adjust the linear combination result. The activation function performs a nonlinear transformation on the result of the linear combination, compressing the value range to (0, 1). Finally, the output of the activation function is converted into a device health score S of (0-100) through linear scaling. This score comprehensively reflects the operating status of the device, and the high and low values ​​reflect the health status of the device. The explainable reasoning unit inputs abnormal states with health scores below a preset threshold into the knowledge graph, matches the fault rule chain corresponding to the abnormal features through a breadth-first search algorithm, and calculates the posterior probability of each fault node in combination with a Bayesian network to generate a reasoning process including the fault path, where the conditional probability table of the Bayesian network is derived based on historical fault data statistics; The diagnostic report generation unit automatically extracts the fault component name, fault type, and characteristic parameter deviation value information based on the reasoning results, calls a predefined report template to generate a structured document, and simultaneously retrieves maintenance plans for similar historical cases from the knowledge graph as maintenance recommendations for the current fault. The report content includes the fault location results, reasoning basis, and maintenance steps, and is transmitted to the human-computer interaction and decision support module via the Ethernet interface.

8. The ship condition monitoring system based on multi-source data according to claim 1, characterized in that: The human-computer interaction and decision support module includes a data visualization unit, an interaction control unit, a historical case management unit and an emergency response support unit; The data visualization unit uses augmented reality technology to construct a digital twin model of the ship using the Unity 3D engine. This model corresponds one-to-one with the actual ship structure and equipment locations. The equipment health score, vibration eigenvector, and operating parameters output by the intelligent fault diagnosis module are then associated with the corresponding components in the digital twin model according to a preset mapping relationship. Abnormal areas are then visualized in the form of a heat map, with color gradients used to represent the degree of abnormality in the equipment status. Finally, WebGL technology is used to enable real-time browser-side rendering of the digital twin model and data visualization results. Crew members can view the status distribution of all equipment on the ship through the terminal interface, obtaining intuitive ship status monitoring information. The interactive control unit integrates touch input and voice recognition modules, where the touch input uses a capacitive touch screen controller that supports multi-touch operation. Then, based on the gesture recognition algorithm, it realizes the control of the monitoring interface through sliding and zooming gestures, as well as the retrieval of detailed data of a single device. Then, it constructs a voice recognition module based on the hidden Markov model, pre-trains a voice command set dedicated to the marine field, collects voice signals through a microphone array and converts them into control commands. At the same time, based on the command parsing engine, the voice commands are mapped to corresponding system operations. Finally, the real-time transmission of interactive commands and system response are achieved through the WebSocket protocol. The historical case management unit is connected to the MySQL database to store structured historical fault data. It can perform keyword retrieval through the Lucene search engine and support multi-dimensional screening by device type and fault type. Staff can compare the feature similarity of the current abnormal state with historical cases to assist in determining the fault mode. The emergency response support unit has a built-in predefined report template library, which automatically extracts fault information from the diagnostic report and fills it into the template to generate a document. At the same time, it retrieves the maintenance knowledge base in the knowledge graph, displays the standard maintenance process corresponding to the fault, and guides maintenance operations.

9. A ship condition monitoring method based on multi-source data, applied to a ship condition monitoring system based on multi-source data according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1. The system collects broadband vibration signals from the ship's bulkhead through a bionic flexible sensing module, while simultaneously using a multi-source data acquisition module to synchronously access data from ship sensors and external data sources. S2. The collected data is processed by the anti-interference communication module and transmitted in a metal shielded and pipeline-dense environment based on metamaterial antennas, frequency hopping spread spectrum technology, and ant colony optimization dynamic routing algorithm; S3. The transmitted multi-source data enters the multi-source data fusion module, which aligns data with different sampling rates through a time synchronization algorithm and uses protocol-independent parsing technology to unify the data format and generate a standardized monitoring data set; S4: The standardized data set is transmitted to the edge computing processing module. A lightweight AI model is used at the cabin edge node to process the data in real time, extracting features of vibration signals and equipment operating parameters, and performing anomaly pre-screening. S5. The results of edge computing processing are input into the intelligent fault diagnosis module, which builds an equipment health model based on multimodal data. By combining knowledge graphs with explainable AI algorithms, the module locates the fault and generates a diagnostic report and maintenance recommendations. S6. The output results of the intelligent fault diagnosis module are presented through the human-computer interaction and decision support module, and the ship status data is visualized in an augmented reality interface, allowing staff to view parameters, compare cases, and obtain emergency response plans.

Images