A ship power equipment fault monitoring method based on a dual-spectrum camera

Abstract

The application relates to a ship power equipment fault monitoring method based on a dual-spectrum camera, which comprises the following steps: acquiring radiation intensity, converting the radiation intensity into temperature values, continuously recording temperature information of each pixel point, and generating infrared imaging temperature data; performing dynamic feature extraction on the infrared imaging temperature data after pretreatment; acquiring spatial information, time information and motion information of a target device through a visible light function of the camera, performing time-frequency domain feature extraction on visual vibration data after pretreatment; aligning the visual vibration data and the infrared imaging temperature data through a sliding window to generate a synchronous time sequence segment; adopting an end-to-end data fusion strategy, processing vibration and temperature multi-modal time sequence data through a deep learning model; fusing an expert knowledge base and an AI model; generating a fault category and a fault probability; and generating a treatment countermeasure in association with a knowledge graph, so that the method has the advantages of improving fault prediction accuracy, improving convenience, improving fault diagnosis reliability and the like.

Description

Technical Field

[0001] This invention relates to the field of ship equipment fault monitoring technology, and in particular to a method for monitoring ship power equipment faults based on a dual-spectrum camera. Background Technology

[0002] As the core of ship operation, the condition monitoring and fault early warning of marine power equipment directly affect navigation safety and operational efficiency. Currently, the mainstream monitoring methods mainly rely on installing contact sensors, such as vibration and temperature sensors, on key parts of the equipment, analyzing physical signals collected along a single path. However, this approach has significant drawbacks: First, contact sensors are complex to install, requiring wiring and fixing on or inside the equipment surface. They are easily damaged and drifted in the harsh environment of high temperature, high humidity, and high vibration on ships, resulting in high maintenance costs and difficulty in achieving comprehensive monitoring of large areas of equipment. Second, vibration and temperature signals are often collected and analyzed by independent systems, resulting in fragmented information that cannot effectively capture the coupled manifestations of faults in multiple physical fields, leading to difficulties in early identification of complex faults and a high false alarm rate. Third, existing intelligent monitoring solutions mostly rely on cloud data analysis or purely data-driven black-box models, which suffer from poor real-time performance and data security risks in scenarios with limited ship network conditions. Furthermore, the model decisions lack interpretability and are difficult to effectively integrate with the experience of domain experts, resulting in insufficient reliability and credibility in practical engineering.

[0003] Therefore, this application provides a method for monitoring ship power equipment faults based on a dual-spectrum camera to solve the above-mentioned problems. Summary of the Invention

[0004] The main objective of this invention is to solve the problems of existing contact sensors, such as complex installation, the need for wiring on the device surface leading to easy damage and drift, the disconnection of vibration and temperature signals, and high false alarm rate.

[0005] This invention provides a method for monitoring faults in marine power equipment based on a dual-spectrum camera, the method comprising:

[0006] The infrared thermal imaging function of the camera detects high-temperature areas of the target device, obtains the radiation intensity, converts the radiation intensity into temperature value, continuously records the temperature information of each pixel, and generates infrared imaging temperature data.

[0007] Dynamic feature extraction is performed after preprocessing the infrared imaging temperature data;

[0008] The spatial, temporal, and motion information of the target device is obtained through the visible light function of the camera. Specifically, this includes: detecting feature points in the initial frame, generating descriptors containing local feature information for the feature points, describing the feature points and tracking their motion trajectory in consecutive frames, and generating visual vibration data.

[0009] Time-frequency domain feature extraction is performed after preprocessing visual vibration data;

[0010] By aligning visual vibration data and infrared imaging temperature data through a sliding window, a synchronized time sequence segment is generated.

[0011] An end-to-end data fusion strategy is adopted, and deep learning models are used to process vibration and temperature multimodal time series data;

[0012] Building an expert knowledge base and integrating it with AI models;

[0013] Fault categories and fault probabilities are generated based on vibration and multimodal time-series data input, AI model fusion, and expert knowledge base.

[0014] By linking knowledge graphs, processing strategies are generated based on vibration and multimodal time-series data.

[0015] Optionally, the step of detecting high-temperature areas of the target device using the infrared thermal imaging function of the camera, obtaining radiation intensity, converting radiation intensity into temperature values, continuously recording the temperature information of each pixel, and generating infrared imaging temperature data specifically includes:

[0016] Cameras calibrated based on ambient temperature compensation and emissivity;

[0017] Thermal imaging images are acquired at a fixed frame rate of 30fps, with each frame containing the radiation intensity of each pixel.

[0018] Convert radiation intensity into temperature information;

[0019] A timestamp is created for each frame of thermal imaging image and stored as a two-dimensional temperature matrix. At the same time, the temperatures of consecutive frames are stored in HDF5 format in chronological order to form infrared imaging temperature data.

[0020] Optionally, the dynamic feature extraction after preprocessing the infrared imaging temperature data includes:

[0021] Denoising is achieved by bilateral filtering and median filtering on each frame of the temperature matrix;

[0022] Normalize the temperature values ​​to the [0,1] range;

[0023] Identify and process mutation values ​​using the Isolation Forest algorithm;

[0024] For the temperature time series data of each pixel, the mean, maximum, minimum and slope are extracted using a sliding window.

[0025] The Sobel operator is used to calculate the gradient of the temperature matrix in the x and y directions for each frame, resulting in a thermal gradient map. Regions with similar temperatures are segmented using a clustering algorithm, and the average temperature, maximum temperature, etc., of each region are calculated. High-temperature regions are marked and their changes over time are tracked.

[0026] Optionally, the step of detecting feature points in the initial frame, generating descriptors containing local feature information for the feature points, describing the feature points and tracking their motion trajectories in consecutive frames, and generating visual vibration data includes:

[0027] In the initial frame, the ORB algorithm is used to detect feature points on the device;

[0028] An ORB descriptor is generated for each feature point, and the ORB descriptor contains local feature information around the feature point.

[0029] The position of feature points in consecutive frames is traced using optical flow.

[0030] Record the displacement of each feature point in consecutive frames to form a motion trajectory;

[0031] By pre-calibrating, the pixel displacement of feature points is converted into actual physical displacement;

[0032] For each feature point, calculate its displacement in the x and y directions, and combine the time interval to obtain the velocity and acceleration;

[0033] The vibration data from multiple feature points are averaged to obtain the overall vibration signal of the equipment.

[0034] Optionally, the time-frequency domain feature extraction after preprocessing the visual vibration data includes:

[0035] Use a high-pass filter to remove the trend term from the vibration signal;

[0036] Wavelet denoising is used to remove high-frequency noise;

[0037] The vibration displacement is normalized to the interval [-1, 1].

[0038] Extract the temporal characteristics of visual vibration data and calculate the peak value, peak-to-peak value, waveform factor and kurtosis of the vibration signal within each sliding window;

[0039] Frequency domain features of visual vibration data are extracted, and the vibration spectrum is calculated by fast Fourier transform to extract the amplitude and dominant frequency of the main frequency components.

[0040] Optionally, aligning visual vibration data and infrared imaging temperature data through a sliding window to generate synchronized time segments includes:

[0041] Ensure that the timestamps of vibration data and temperature data are synchronized;

[0042] Linear interpolation is performed on the vibration data to make its sampling rate consistent with that of the temperature data;

[0043] Align vibration and temperature data using a sliding window;

[0044] Generate a fused data block containing temperature matrix and vibration data, i.e., a synchronization time segment.

[0045] Optionally, the end-to-end data fusion strategy, which processes vibration and temperature multimodal time-series data through a deep learning model, includes:

[0046] A dual-stream Transformer model is adopted. 1D-CNN extracts local features from vibration signals, LSTM captures temporal dependencies and outputs vibration feature vectors, 3D-CNN processes the temperature matrix and extracts spatiotemporal features, and the spatial attention module focuses on the high-temperature region and outputs temperature feature vectors.

[0047] By dynamically weighting and fusing dual-modal features through a gated cross-attention mechanism, the weights of vibration and temperature features are automatically learned.

[0048] Use a labeled dataset to train and deploy the model.

[0049] Optionally, the construction of the expert knowledge base and the integration of the AI ​​model includes:

[0050] Collect information on equipment failure cases, maintenance manuals, and expert experience to build a knowledge graph.

[0051] Manage IF-THEN rules using the rule engine Drools;

[0052] Store the knowledge graph using Neo4j;

[0053] Incorporate expert specifications as constraints into the model training process.

[0054] Optionally, the step of generating fault categories and fault probabilities based on vibration and multimodal time-series data input, AI model fusion, and expert knowledge base includes:

[0055] The AI ​​model takes a fused feature vector of vibration and temperature as input and outputs the probability of each fault category.

[0056] The rules engine checks whether the expert rules are met and corrects the model output accordingly;

[0057] An alarm is triggered based on the set threshold, and the current fault category is determined.

[0058] Optionally, the associated knowledge graph, which generates processing strategies based on vibration and multimodal time-series data, includes: querying the corresponding processing strategies in the knowledge graph according to the predicted fault category and displaying them on the system interface, writing the processing results into the knowledge base, and updating the model weights through reinforcement learning.

[0059] This application provides a method for monitoring faults in marine power equipment based on a dual-spectrum camera. Compared with traditional contact-based temperature and vibration monitoring, this method collects and fuses data from multiple distances using a single camera, improving the accuracy of fault prediction. In addition, non-contact vibration and temperature detection enhances the convenience of monitoring. Furthermore, this application improves the reliability of fault diagnosis by using an AI model and integrating an expert knowledge base. Attached Figure Description

[0060] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0061] This invention provides a method for fault monitoring of marine power equipment based on a dual-spectrum camera, comprising: detecting high-temperature areas of the target equipment using the infrared thermal imaging function of the camera, acquiring radiation intensity, converting the radiation intensity into temperature values, continuously recording the temperature information of each pixel, and generating infrared imaging temperature data; performing dynamic feature extraction after preprocessing the infrared imaging temperature data; acquiring spatial, temporal, and motion information of the target equipment using the visible light function of the camera, specifically including: detecting feature points in the initial frame, generating descriptors containing local feature information for the feature points, describing the feature points and tracking their motion trajectories in consecutive frames, and generating visual vibration data; and preprocessing the visual vibration data. Then, time-frequency domain feature extraction is performed; visual vibration data and infrared imaging temperature data are aligned through a sliding window to generate synchronous time-series segments; an end-to-end data fusion strategy is adopted, and vibration and temperature multimodal time-series data are processed through a deep learning model; an expert knowledge base is constructed and fused with the AI ​​model; based on the input vibration and multimodal time-series data, AI model fusion, and expert knowledge base, fault categories and fault probabilities are generated; a knowledge graph is associated to generate processing countermeasures based on vibration and multimodal time-series data; the main purpose of this invention is to solve the problems of complex installation, easy damage and drift due to wiring on the device surface, disconnection of vibration and temperature signals, and high false alarm rate of existing contact sensors.

[0062] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0063] Example

[0064] Reference manual attached Figure 1 This embodiment provides a method for monitoring ship propulsion equipment faults based on a dual-spectrum camera. The method includes:

[0065] The infrared thermal imaging function of the camera detects high-temperature areas of the target device, obtains the radiation intensity, converts the radiation intensity into temperature value, continuously records the temperature information of each pixel, and generates infrared imaging temperature data.

[0066] Dynamic feature extraction is performed after preprocessing the infrared imaging temperature data;

[0067] The spatial, temporal, and motion information of the target device is obtained through the visible light function of the camera. Specifically, this includes: detecting feature points in the initial frame, generating descriptors containing local feature information for the feature points, describing the feature points and tracking their motion trajectory in consecutive frames, and generating visual vibration data.

[0068] Time-frequency domain feature extraction is performed after preprocessing visual vibration data;

[0069] By aligning visual vibration data and infrared imaging temperature data through a sliding window, a synchronized time sequence segment is generated.

[0070] An end-to-end data fusion strategy is adopted, and deep learning models are used to process vibration and temperature multimodal time series data;

[0071] Building an expert knowledge base and integrating it with AI models;

[0072] Fault categories and fault probabilities are generated based on vibration and multimodal time-series data input, AI model fusion, and expert knowledge base.

[0073] By linking knowledge graphs, processing strategies are generated based on vibration and multimodal time-series data.

[0074] The infrared thermal imaging function of the camera detects high-temperature areas of the target device, acquires the radiation intensity, converts the radiation intensity into a temperature value, and continuously records the temperature information of each pixel to generate infrared imaging temperature data. Specifically, this includes:

[0075] Cameras calibrated based on ambient temperature compensation and emissivity;

[0076] Thermal imaging images are acquired at a fixed frame rate of 30fps, with each frame containing the radiation intensity of each pixel.

[0077] Convert radiation intensity into temperature information;

[0078] A timestamp is created for each frame of thermal imaging image and stored as a two-dimensional temperature matrix. At the same time, the temperatures of consecutive frames are stored in HDF5 format in chronological order to form infrared imaging temperature data.

[0079] Dynamic feature extraction after preprocessing infrared imaging temperature data includes:

[0080] Denoising is achieved by bilateral filtering and median filtering on each frame of the temperature matrix;

[0081] Normalize the temperature values ​​to the [0,1] range;

[0082] Identify and process mutation values ​​using the Isolation Forest algorithm;

[0083] For the temperature time series data of each pixel, the mean, maximum, minimum and slope are extracted using a sliding window.

[0084] The Sobel operator is used to calculate the gradient of the temperature matrix in the x and y directions for each frame, resulting in a thermal gradient map. Regions with similar temperatures are segmented using a clustering algorithm, and the average temperature, maximum temperature, etc., of each region are calculated. High-temperature regions are marked and their changes over time are tracked.

[0085] Feature points are detected in the initial frame, descriptors containing local feature information are generated for each feature point, and the feature points are described and their motion trajectories are tracked in consecutive frames to generate visual vibration data, including:

[0086] In the initial frame, the ORB algorithm is used to detect feature points on the device; feature points can be selected from device edges or bolts.

[0087] An ORB descriptor is generated for each feature point. The ORB descriptor contains local feature information around the feature point.

[0088] The position of feature points in consecutive frames is traced using optical flow.

[0089] Record the displacement of each feature point in consecutive frames to form a motion trajectory;

[0090] By pre-calibrating, the pixel displacement of feature points is converted into actual physical displacement;

[0091] For each feature point, calculate its displacement in the x and y directions, and combine the time interval to obtain the velocity and acceleration;

[0092] The vibration data from multiple feature points are averaged to obtain the overall vibration signal of the equipment.

[0093] The time-frequency domain feature extraction after preprocessing visual vibration data includes:

[0094] Use a high-pass filter to remove the trend term from the vibration signal;

[0095] Wavelet denoising is used to remove high-frequency noise;

[0096] The vibration displacement is normalized to the interval [-1, 1].

[0097] Extract the temporal characteristics of visual vibration data and calculate the peak value, peak-to-peak value, waveform factor and kurtosis of the vibration signal within each sliding window;

[0098] Frequency domain features of visual vibration data are extracted, and the vibration spectrum is calculated by fast Fourier transform to extract the amplitude and dominant frequency of the main frequency components.

[0099] By aligning visual vibration data and infrared imaging temperature data through a sliding window, synchronized time segments are generated, including:

[0100] Ensure that the timestamps of vibration data and temperature data are synchronized;

[0101] Linear interpolation is performed on the vibration data to make its sampling rate consistent with that of the temperature data;

[0102] Vibration and temperature data are aligned using a sliding window; the window size of the sliding window can be 10 seconds, and the step size of the sliding window is 1 second.

[0103] Generate a fused data block containing temperature matrix and vibration data, i.e., a synchronous time segment. The temperature matrix includes spatial and temporal data, and the vibration data is temporal data.

[0104] An end-to-end data fusion strategy is adopted, and deep learning models are used to process vibration and temperature multimodal time series data, including:

[0105] A dual-stream Transformer model is adopted. 1D-CNN extracts local features from vibration signals, LSTM captures temporal dependencies and outputs vibration feature vectors, 3D-CNN processes the temperature matrix and extracts spatiotemporal features, and the spatial attention module focuses on the high-temperature region and outputs temperature feature vectors.

[0106] By dynamically weighting and fusing dual-modal features through a gated cross-attention mechanism, the weights of vibration and temperature features are automatically learned.

[0107] Use a labeled dataset to train and deploy the model.

[0108] Building an expert knowledge base and integrating it with AI models includes:

[0109] Collect information on equipment failure cases, maintenance manuals, and expert experience to build a knowledge graph.

[0110] Manage IF-THEN rules using the rule engine Drools;

[0111] Store the knowledge graph using Neo4j;

[0112] Incorporate expert specifications as constraints into the model training process.

[0113] Based on vibration and multimodal time-series data input, AI model fusion, and expert knowledge base, fault categories and fault probabilities are generated, including:

[0114] The AI ​​model takes a fused feature vector of vibration and temperature as input and outputs the probability of each fault category.

[0115] The rules engine checks whether the expert rules are met and corrects the model output accordingly;

[0116] An alarm is triggered based on a set threshold, and the current fault category is determined. The threshold can be set to 0.8, for example, triggering an alarm when the highest probability is greater than 0.8.

[0117] By associating with the knowledge graph, processing strategies are generated based on vibration and multimodal time-series data. This includes: querying the corresponding processing strategies in the knowledge graph based on the predicted fault category and displaying them on the system interface; writing the processing results into the knowledge base; and updating the model weights through reinforcement learning.

[0118] In a specific embodiment, the camera's visible light function detects abnormal vibration of the main unit's supercharger (increased vibration velocity and kurtosis); the camera's infrared thermal imaging function detects an increase in temperature in the corresponding area of ​​the supercharger bearing; as time progresses, the vibration frequency exhibits bearing fault characteristic frequencies, the temperature continues to rise, and the area expands; the model's vibration branch (1D-CNN) extracts vibration waveform features and identifies a periodic degradation trend; the model's temperature branch (3D-CNN) captures the direction of heat diffusion and focuses on high-temperature areas; the model's gated cross-attention outputs vibration feature weights, temperature feature weights, and fusion confidence; the system calculates and outputs fault categories and probabilities, such as bearing wear, blade carbon buildup, and impeller imbalance; rules are triggered and alarms are activated, determining that the bearing wear probability is the highest and exceeds a set threshold based on vibration and temperature features; the system provides handling countermeasures and suggestions based on the fault category; the knowledge base is updated, and case data is added to the training set.

[0119] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring faults in marine power equipment based on a dual-spectrum camera, characterized in that, include: The infrared thermal imaging function of the camera detects high-temperature areas of the target device, obtains radiation intensity, converts radiation intensity into temperature value, continuously records the temperature information of each pixel, and generates infrared imaging temperature data. Dynamic feature extraction is performed after preprocessing the infrared imaging temperature data; Acquiring spatial, temporal, and motion information of a target device using the visible light function of a camera specifically includes: detecting feature points in an initial frame, generating descriptors containing local feature information for the feature points, describing the feature points and tracking their motion trajectories in consecutive frames, and generating visual vibration data; wherein, the step of detecting feature points in the initial frame, generating descriptors containing local feature information for the feature points, describing the feature points and tracking their motion trajectories in consecutive frames, and generating visual vibration data includes: In the initial frame, the ORB algorithm is used to detect feature points on the device; An ORB descriptor is generated for each feature point, and the ORB descriptor contains local feature information around the feature point. The position of feature points in consecutive frames is traced using optical flow. Record the displacement of each feature point in consecutive frames to form a motion trajectory; By pre-calibrating, the pixel displacement of feature points is converted into actual physical displacement; For each feature point, calculate its displacement in the x and y directions, and combine the time interval to obtain the velocity and acceleration; The vibration data from multiple feature points are averaged to obtain the overall vibration signal of the equipment. Time-frequency domain feature extraction is performed after preprocessing visual vibration data; By aligning visual vibration data and infrared imaging temperature data through a sliding window, a synchronized time sequence segment is generated. An end-to-end data fusion strategy is adopted to process vibration and temperature multimodal time-series data through a deep learning model; wherein, the end-to-end data fusion strategy and the processing of vibration and temperature multimodal time-series data through a deep learning model include: A dual-stream Transformer model is adopted. 1D-CNN extracts local features from vibration signals, LSTM captures temporal dependencies and outputs vibration feature vectors, 3D-CNN processes the temperature matrix and extracts spatiotemporal features, and the spatial attention module focuses on the high-temperature region and outputs temperature feature vectors. By dynamically weighting and fusing dual-modal features through a gated cross-attention mechanism, the weights of vibration and temperature features are automatically learned. Use labeled datasets to train and deploy the model; Building an expert knowledge base and integrating it with AI models; Fault categories and fault probabilities are generated based on vibration and multimodal time-series data input, AI model fusion, and expert knowledge base. By linking knowledge graphs, processing strategies are generated based on vibration and multimodal time-series data.

2. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 1, characterized in that, The process of detecting high-temperature areas of the target device using the infrared thermal imaging function of the camera, acquiring radiation intensity, converting radiation intensity into temperature values, continuously recording the temperature information of each pixel, and generating infrared imaging temperature data specifically includes: Cameras calibrated based on ambient temperature compensation and emissivity; Thermal imaging images are acquired at a fixed frame rate of 30fps, with each frame containing the radiation intensity of each pixel. Convert radiation intensity into temperature information; A timestamp is created for each frame of thermal imaging image and stored as a two-dimensional temperature matrix. At the same time, the temperatures of consecutive frames are stored in HDF5 format in chronological order to form infrared imaging temperature data.

3. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 2, characterized in that, The dynamic feature extraction after preprocessing the infrared imaging temperature data includes: Denoising is achieved by bilateral filtering and median filtering on each frame of the temperature matrix; Normalize the temperature values ​​to the [0,1] range; Identify and process mutation values ​​using the Isolation Forest algorithm; For the temperature time series data of each pixel, the mean, maximum, minimum and slope are extracted using a sliding window. The Sobel operator is used to calculate the gradient of the temperature matrix in the x and y directions for each frame, resulting in a thermal gradient map. Regions with similar temperatures are segmented using a clustering algorithm, and the average temperature, maximum temperature, etc., of each region are calculated. High-temperature regions are marked and their changes over time are tracked.

4. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 3, characterized in that, The time-frequency domain feature extraction after preprocessing the visual vibration data includes: Use a high-pass filter to remove the trend term from the vibration signal; Wavelet denoising is used to remove high-frequency noise; The vibration displacement is normalized to the interval [-1, 1]. Extract the temporal characteristics of visual vibration data and calculate the peak value, peak-to-peak value, waveform factor and kurtosis of the vibration signal within each sliding window; Frequency domain features of visual vibration data are extracted, and the vibration spectrum is calculated by fast Fourier transform to extract the amplitude and dominant frequency of the main frequency components.

5. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 4, characterized in that, The step of aligning visual vibration data and infrared imaging temperature data through a sliding window to generate a synchronized time segment includes: Ensure that the timestamps of vibration data and temperature data are synchronized; Linear interpolation is performed on the vibration data to make its sampling rate consistent with that of the temperature data; Align vibration and temperature data using a sliding window; Generate a fused data block containing temperature matrix and vibration data, i.e., a synchronization time segment.

6. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 5, characterized in that, The integration of expert knowledge base and AI model construction includes: Collect equipment failure case information, maintenance manual information, and expert experience information to construct a knowledge graph; Manage IF-THEN rules using the rule engine Drools; Store the knowledge graph using Neo4j; Incorporate expert specifications as constraints into the model training process.

7. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 6, characterized in that, The generation of fault categories and fault probabilities based on vibration and multimodal time-series data input, AI model fusion, and expert knowledge base includes: The AI ​​model takes a fused feature vector of vibration and temperature as input and outputs the probability of each fault category. The rules engine checks whether the expert rules are met and corrects the model output accordingly; An alarm is triggered based on the set threshold, and the current fault category is determined.

8. The method for monitoring ship propulsion equipment faults based on a dual-spectrum camera according to claim 7, characterized in that, The associated knowledge graph generates processing strategies based on vibration and multimodal time-series data, including: querying the corresponding processing strategies in the knowledge graph according to the predicted fault category and displaying them on the system interface; writing the processing results into the knowledge base; and updating the model weights through reinforcement learning.

Images