Ship power system abnormity monitoring method based on multi-modal pre-training large model

Abstract

The invention provides a ship power system abnormity monitoring method, which integrates multi-source and multi-mode monitoring state information of a ship power system, and realizes high-efficiency monitoring of ship power system equipment and system operation states. A ship power system normal / abnormal sample is simulated and generated through a three-dimensional model, a simulation model and a signal processing transformation technology by utilizing an existing abnormal sample and expert knowledge, and a CLIP large model is finely adjusted by an intelligent computing platform based on cloud-edge collaboration, so that a scheme for monitoring the abnormity of the ship power system by using a small amount of abnormal data is realized; according to the method, the binding of the picture and the text is realized through the CLIP model, namely, the CLIP can obtain the exception description in the picture, and a subsequent processing scheme can be generated through the exception description, so that the accurate pushing of the exception alarm is realized, and useless exception alarm signals are shielded for irrelevant operators.

Description

Technical Field

[0001] The present invention belongs to the technical field of ship power system anomaly detection, and in particular relates to a ship power system anomaly monitoring method based on a multimodal pre-trained large model. Background Art

[0002] The power system is the most core system in the entire ship. Its safety and reliability directly affect the safe navigation of the ship. With the development of automated control and intelligent technology, the automation level and complexity of the ship's power system have gradually improved. The demand for less-manned and unmanned inspections of ships has gradually increased, which has put forward higher requirements for the reliability and safety of the power system, and has paid more and more attention to the abnormal monitoring of the power system.

[0003] Currently, due to the limited number of abnormal samples in ship power systems, small models based on end-to-end deep learning neural networks generally generalize poorly, and the application of conventional artificial intelligence algorithms is limited. Secondly, most methods for fault diagnosis and abnormality monitoring of ship power systems focus on specific equipment such as turbines, diesel engines, pumps, and valves, and fail to comprehensively analyze fault diagnosis and abnormality monitoring for the entire power system. Furthermore, with the implementation of large-scale abnormality monitoring for power system status, the number of system abnormality alarms will increase significantly. Excessive alarms will seriously affect operator control. Therefore, it is necessary to filter out useless alarm signals and provide effective alarm signals based on the responsibilities of ship power system operators.

[0004] Existing abnormality monitoring schemes for ship power systems are usually carried out for key power equipment. There is insufficient fusion of various status monitoring information, insufficient systematic design, and insufficient overall monitoring efficiency. The abnormality monitoring algorithm relies on a large amount of real abnormal data. At the same time, the increase in the amount of status monitoring leads to an increase in abnormal alarms, which affects the normal operation of the ship power system operator. Summary of the Invention

[0005] In order to solve the technical problems existing in the prior art, the present invention provides a method for monitoring abnormalities in a ship power system, comprising:

[0006] Multi-source and multi-modal status information fusion monitoring: Based on the ship's power system, sensors are deployed at key locations to monitor the status of corresponding locations. The sensor monitoring information is converted into images, and the image information is processed through a multi-modal large model to determine whether the thermal status of the system is normal;

[0007] A small number of sample anomaly monitoring methods are designed, which use a multimodal pre-trained large model to realize system anomaly monitoring. The existing three-dimensional model of the ship power system and various working condition simulation models are used, combined with a small amount of real abnormal data and expert knowledge and experience, to generate normal and abnormal samples.

[0008] Furthermore, multi-modal state information fusion monitoring specifically includes:

[0009] Based on the layout of the equipment and connecting pipelines in the ship's power system engine room, key areas that need to be monitored are analyzed. Infrared cameras, sound acquisition sensors, and vibration monitoring sensors are installed at appropriate locations on the system or equipment to monitor the status of the power system's mechanical equipment, pipelines, etc. in real time. The sound and vibration signals are converted into images through signal conversion methods such as Mel spectrum conversion. The video and image information after sound and vibration conversion are processed through a large multimodal model. The video, sound, and vibration information are integrated to comprehensively judge whether the power system's mechanical equipment and connecting pipelines have any abnormal conditions such as mechanical damage or deformation;

[0010] Temperature, pressure, flow and other monitoring sensors are arranged at key locations on the power system pipelines to realize real-time monitoring of the system's thermal state. The temperature, pressure and flow under the system's operating state are integrated into image information that can be processed by a large multi-modal model to comprehensively judge whether the system's thermal state is normal.

[0011] Furthermore, the design of the small sample anomaly detection method includes:

[0012] Generate normal samples of mechanical equipment's appearance and structure, as well as abnormal samples such as loosening and deformation of key screws;

[0013] Generate normal and abnormal samples of rotating machinery vibration signals;

[0014] Normal and abnormal sample generation of system operation status;

[0015] The three methods mentioned above are used to generate normal and abnormal samples of the ship power system and fine-tune the CLIP model.

[0016] Furthermore, the normal and abnormal data generated for loosening and deformation of key screws in mechanical equipment specifically include:

[0017] Capture multi-angle images of equipment and pipelines in the area that needs to be monitored in the 3D model of the ship's power system, and add descriptions to the images;

[0018] Based on existing images of pipeline deformation and loosening of key screws of mechanical equipment or expert knowledge, abnormal conditions of mechanical equipment such as pipeline deformation and loosening of key screws are simulated in a three-dimensional model, and images are captured as abnormal data samples.

[0019] Furthermore, the generation of normal and abnormal samples of rotating machinery vibration signals specifically includes:

[0020] Vibration simulation analysis of rotating machinery is used to simulate and generate normal and abnormal vibration signals. Vibration frequency data is converted into Mel spectrum feature time-frequency graphs through Mel spectrum feature transformation, and the vibration frequency time series signal is converted into images.

[0021] Based on existing rotating machinery vibration signals or expert knowledge, a Mel-spectrum feature time-frequency diagram of the abnormal equipment vibration frequency is generated and labeled "abnormal vibration of rotating machinery equipment", forming a series of image-text pairs.

[0022] Furthermore, the generation of normal and abnormal system operation status samples specifically includes:

[0023] Convert temperature, pressure, and flow values ​​to 0-255 and convert them into images to suit CLIP model processing;

[0024] The temperature, pressure and flow values ​​at various locations in the system are produced by the power system simulation model under normal and abnormal conditions;

[0025] By integrating the above methods, "main power system heat dissipation is normal" is marked on the normal state value, and "main power system heat dissipation is abnormal" is marked on the abnormal state value.

[0026] Furthermore, normal and abnormal samples of the ship power system are generated, and the CLIP model is fine-tuned including:

[0027] Convert the description of the anomaly into English and use it as the input of the CLIP model. Use the encoder to convert the image and text into the same space for comparison.

[0028] The CLIP model optimizes the symmetric cross-entropy loss so that correct image and text pairs are closer in space and incorrect pairs are further away.

[0029] Furthermore, it also includes:

[0030] The intelligent alarm method is designed. After processing the image through the CLIP model, the probability value of the similarity between the image and the abnormal description can be obtained. The abnormal description with the greatest similarity is selected as the abnormal information in the image. That is, the abnormal information in the image can be obtained through the CLIP model.

[0031] Based on the abnormality description, the handling method can be preset through expert knowledge and experience, and the relevant personnel of the abnormality can be listed according to the ship power system operating procedures;

[0032] Based on the list of relevant personnel of the exception, intelligent push is implemented to operators who need to participate in the processing or need to know the exception.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention integrates multi-source and multi-modal monitoring status information of the ship power system to realize high-efficiency monitoring of the ship power system equipment and system operation status; utilizes existing abnormal samples and expert knowledge, simulates and generates normal / abnormal samples of the ship power system through three-dimensional models, simulation models and signal processing transformation technology, and fine-tunes the CLIP large model based on the cloud-edge collaborative intelligent computing platform, realizing a solution for abnormal monitoring of the ship power system using a small amount of abnormal data; through the CLIP model, the binding of pictures and texts is realized, that is, CLIP obtains the abnormal description in the picture, and the subsequent processing plan can be generated based on the abnormal description, realizing the accurate push of abnormal alarms and shielding useless abnormal alarm signals for irrelevant operators. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram for data fusion and sample generation of ship power systems;

[0036] Figure 2 Fine-tune the training graph for the CLIP model;

[0037] Figure 3 This is a diagram showing the implementation plan for the ship power system anomaly monitoring algorithm based on a multimodal pre-trained large model in cloud-edge collaboration. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0039] The present invention provides a method for monitoring abnormalities in a ship power system, comprising:

[0040] 1) Multi-source and multi-modal state information fusion monitoring

[0041] Based on the layout of the equipment and connecting pipes within the ship's power system engine room, key areas requiring monitoring are analyzed. Infrared cameras, sound acquisition sensors, and vibration monitoring sensors are installed at appropriate locations on the system or equipment to monitor the status of the power system's mechanical equipment in real time. Signal transformation methods such as Mel spectrum transform are used to convert sound and vibration signals into images. A multimodal large model is used to process the video and converted image information. The video, sound, and vibration information are integrated to comprehensively determine whether the power system's mechanical equipment and connecting pipes are experiencing mechanical damage, deformation, or other abnormal conditions. Temperature, pressure, and flow monitoring sensors are placed at key locations on the power system's pipelines to monitor the system's thermal status in real time. The temperature, pressure, and flow rates during system operation are integrated into image information that can be processed by the multimodal large model to comprehensively determine whether the system's thermal status is normal.

[0042] 2) Design of small sample anomaly detection method

[0043] System anomaly detection is achieved using a large multimodal pre-trained model (CLIP). The CLIP model learns to recognize images by simultaneously interpreting images and text. The CLIP model's pre-training task is to predict which image matches which text. It is pre-trained using a dataset of 400 million (image, text) pairs collected from the internet. After training, CLIP can use natural language to identify and describe objects it has never seen before. Without requiring specific task training, it can perform well on a wide range of tasks and demonstrates strong learning and generalization capabilities.

[0044] Traditional methods for monitoring abnormalities in ship power systems usually require a large amount of real abnormal data from ship power machinery and equipment. However, in most cases, there are not enough real abnormal data samples available. The CLIP model has demonstrated strong generalization capabilities in tasks without training samples in specific fields, and can play an important role in the field of abnormality monitoring where training samples are lacking. The CLIP model itself recognizes the similarity between images and texts. In order to make it suitable for normal and abnormal monitoring, it is necessary to match the images of the engine room in normal conditions with the normal descriptions, and the images in abnormal conditions with the abnormal descriptions. To this end, it is necessary to use the abnormal data of the power system to fine-tune the CLIP model. However, in the case that the number of available abnormal samples of the ship power system is very limited, the present invention combines the development of current digital technology. During the design process of the ship power system, a real and detailed three-dimensional model of the ship power system is drawn, and an operation simulation model of the power system under various working conditions is also constructed. A method for generating normal and abnormal samples by utilizing the existing three-dimensional model of the ship power system and the simulation models of various working conditions, combined with a small amount of real abnormal data and expert knowledge and experience, is proposed. The specific method is as follows:

[0045] (1) Generate normal / abnormal data such as loosening and deformation of key screws of mechanical equipment

[0046] Within the 3D model of the ship's power system, capture multi-angle images of equipment and pipelines in the area requiring monitoring and add descriptions to the images. Based on existing images of pipeline deformation and loosening of key screws in mechanical equipment, or based on expert knowledge, simulate abnormal mechanical equipment conditions such as pipeline deformation and loosening of key screws in the 3D model and capture these images as abnormal data samples. For example, if a picture of a loosening key screw in a critical power machinery component affecting system safety is captured, the image description could be annotated as "Loosening of key screw in a power machinery component."

[0047] (2) Generating normal / abnormal samples of rotating machinery vibration signals

[0048] Vibration simulation analysis of rotating machinery is used to simulate and generate normal / abnormal vibration signals. Based on the fact that experienced personnel in ship power systems can determine whether equipment such as steam turbines are faulty by "listening to the rod," this invention converts vibration frequency data into Mel-spectrogram features (MFCCs). MFCC features can map linear representations (such as short-time Fourier transforms) to a logarithmic scale to mimic human auditory responses. The Mel-spectrogram feature transformation converts the vibration frequency time series signal into an image. Based on existing rotating machinery vibration signals or expert knowledge, MFCCs are generated for normal / abnormal equipment vibration frequencies. Normal state values ​​are labeled "abnormal vibration of rotating machinery equipment," and abnormal state values ​​are labeled "abnormal vibration of rotating machinery equipment." This creates a series of (image, text) pairs used to train the CLIP model to learn the normal / abnormal states of ship power systems.

[0049] (3) Generate samples of normal / abnormal system operation status

[0050] System operating state variables are typically temperature, pressure, and flow. To integrate temperature, pressure, and flow information with ambient temperature and pressure information, and to adapt to CLIP model processing, the temperature, pressure, and flow signals are fused into an image. From a temporal and spatial perspective, the temperature and pressure are combined with the ambient temperature and pressure near the power system pipelines. The flow is expanded to the same dimensions as the temperature and pressure through replication, allowing the three two-dimensional plane variables of temperature, pressure, and flow to be stacked and fused. To this end, the temperature, pressure, and flow values ​​are converted to 0-255 and converted into an image, suitable for CLIP model processing. The power system simulation model is used to generate temperature, pressure, and flow values ​​at various points in the system under normal and abnormal conditions. These values ​​are fused using the above method, with the normal state values ​​labeled "main power system heat removal normal" and the abnormal state values ​​labeled "main power system heat removal abnormal."

[0051] The above three methods are used to generate normal / sample of ship power system (see Appendix Figure 1 ) fine-tuning the CLIP model, and its training process (see Appendix Figure 2 ) are as follows:

[0052] Because the open-source CLIP model is trained based on English text, the description of the anomaly needs to be converted into English as input to the CLIP model. An encoder is used to convert the image and text into the same space for comparison. The CLIP model optimizes the symmetric cross-entropy loss to make correct image and text pairs closer in space and incorrect pairs farther apart.

[0053] 3) Intelligent alarm method design

[0054] After processing the image through the CLIP model, the probability value of the similarity between the image and the abnormal description can be obtained. The abnormal description with the greatest similarity is selected as the abnormal information in the image. That is, the abnormal information in the image can be obtained through the CLIP model. Based on the abnormal description, the processing method can be preset based on expert knowledge and experience. According to the ship power system operating procedures, the relevant personnel of the abnormality can be listed. Based on the list of relevant personnel of the abnormality, intelligent push is realized to the operators who need to participate in the processing or need to know the abnormality.

[0055] The architecture of the present invention is the "cloud-edge-end" architecture. Pressure, temperature, flow, vibration, voiceprint, video and other sensors are deployed on the field side to obtain on-site system and equipment status information; computing power is deployed on the edge side to complete data preprocessing; and a high-efficiency intelligent computing platform is deployed on the cloud to complete the fine-tuning and inference operations of the multimodal pre-trained large model.

[0056] (1) Analyze key monitoring areas and generate a summary table of abnormal descriptions

[0057] The present invention comprehensively considers the abnormal monitoring of the ship power system and realizes full-range monitoring of the physical state and operating state of the power system equipment. First, according to the equipment layout, pipeline layout, equipment composition, and electrical cable layout of the ship power system, a camera with infrared function is set to realize no-dead-angle monitoring of the power engine room equipment, pipelines, and electrical cable connections. Sound and vibration monitoring sensors are arranged on the rotating mechanical equipment body, and sound and vibration monitoring sensors are arranged on the system main pipeline to realize sound and vibration monitoring of the system pipelines. Temperature, pressure, and flow sensors are arranged at the entrances and exits of the main equipment of the power system to monitor the operating status of the system. Based on the existing experience of ship power systems, the common abnormal types of ship power systems are sorted out in the following table:

[0058]

[0059]

[0060] (2) State monitoring information fusion design

[0061] Abnormal information about a ship's power system and equipment can be reflected through multi-source, multi-modal information such as video, sound, and vibration signals. Using signal transformation methods such as Mel-spectrogram transform, sound and vibration signals are converted into images. A large multimodal model processes the converted video and sound and vibration images. The video, sound, and vibration status information is integrated to comprehensively determine whether the power system's mechanical equipment and connecting piping are experiencing abnormal conditions such as mechanical damage or deformation. The temperature, pressure, and flow rate of a ship's power system's thermal operating state are correlated. By combining the system's operating temperature, pressure, and flow rate information with the ambient temperature and pressure information to create a holistic image suitable for CLIP model processing, the integration of ship power system operating status information is achieved.

[0062] (3) Generate normal / abnormal samples

[0063] The existing three-dimensional model of the ship power system is used, and an automatic image capture script is set up to generate normal state images, forming (normal image, normal description) pairs as training samples; based on the existing expert knowledge of loose screws, loose cable plugs, etc. or abnormalities, key abnormalities such as loose screw faults, loose or falling cable plugs, pipeline deformation, and leakage at pipeline or flange connections are set, and a simulation injection script is written to inject different abnormalities into different positions in the three-dimensional model of the power system, and automatically take screenshots to form (abnormal image, abnormal description) pairs with the injected abnormal patterns as training samples.

[0064] The vibration simulation analysis model of rotating machinery equipment is used to generate normal / abnormal vibration frequency time series diagrams of rotating machinery equipment. The vibration frequency time series diagrams are then transformed by MFCC to generate normal / abnormal time-frequency diagrams. The pairs of (normal vibration frequency time-frequency diagram, normal equipment vibration) and (abnormal vibration frequency time-frequency diagram, abnormal equipment vibration) are formed as training samples.

[0065] The ship power system simulation model is used to generate temperature, pressure, and flow curves of the inlet and outlet of each major equipment under different working conditions. Combined with the temperature and pressure in the environment near the power system pipeline, the temperature and pressure are expanded into a two-dimensional plane. A time window is set to intercept the temperature, pressure, and flow curves. The flow is expanded to the same dimension as the temperature and pressure by copying. At this time, the temperature, pressure, and flow can be expanded into a two-dimensional plane. The two-dimensional temperature, pressure, and flow three-variable surfaces are stacked and fused. The temperature, pressure, and flow values ​​are converted to 0-255, that is, converted into images corresponding to RGB, forming (normal curve of main power system heat removal, normal main power system heat removal) pairs and (normal curve of auxiliary system heat removal, normal auxiliary system heat removal) pairs as training samples. In the power system simulation model, the main power system heat dissipation anomaly and the auxiliary system heat dissipation anomaly are injected under different working conditions. The temperature, pressure and flow curves under the abnormal state are fused into a picture according to the method of fusing the temperature, pressure and flow curves into a picture under normal operating conditions, forming (control valve stuck abnormality curve graph, control valve stuck abnormality) pair, (switch valve state error fault curve graph, switch valve state error fault) pair, (main power system heat dissipation anomaly curve graph, main power system heat dissipation anomaly) pair, (auxiliary system operation state abnormality curve graph, auxiliary system operation state abnormality) pair as training samples.

[0066] (4) CLIP model fine-tuning training and inference

[0067] See the attached diagram for specific training and reasoning Figure 2 Because the CLIP model is pre-trained in English, the description of the anomaly is converted into an English description as the input of the CLIP model. The encoder is used to convert the image and text into a vector of a common dimension so that they can be compared in the same space. The CLIP model optimizes the symmetric cross entropy loss so that the correct image and text pairs are closer in space and the incorrect pairs are farther away in space.

[0068] A cloud-edge collaborative high-performance computing service platform is set up as an intelligent computing platform for ship power system anomaly monitoring to carry out CLIP model fine-tuning training and reasoning; in order to optimize the transmission of ship power system monitoring information, a cloud-edge collaborative computing technology is used to build a cloud-edge collaborative power system intelligent monitoring platform (see Appendix Figure 3 Deploy edge video recognition and monitoring devices on-site in the power engine compartment to process image data, edge audio information recognition and monitoring devices to process vibration frequency data, and edge system operation status information recognition and monitoring devices to process power system operation status data. This brings computing power down to the on-site equipment and improves the efficiency of power system operation status information transmission. Fine-tune training and real-time inference of the CLIP model are performed in the cloud.

[0069] (5) Intelligent abnormal alarm

[0070] If a ship's power system anomaly occurs, the CLIP model processes the anomaly image and outputs a description of the anomaly. Based on expert knowledge and experience, it forms an anomaly analysis and processing tree, designs a corresponding anomaly handling plan, lists the personnel responsible for the anomaly, and intelligently pushes the information to relevant personnel.

[0071] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent changes and modifications made according to the patent scope of the present invention and the contents of the specification are within the scope covered by the patent of the present invention.

Claims

1. A method for monitoring abnormality of a ship power system, characterized in that: include: Multi-source and multi-modal status information fusion monitoring: Based on the ship's power system, sensors are deployed at key locations to monitor the status of corresponding locations. The sensor monitoring information is converted into images, and the image information is processed through a multi-modal large model to determine whether the system is operating normally. A small number of sample anomaly monitoring methods are designed, which use a multimodal pre-trained large model to realize system anomaly monitoring. The existing three-dimensional model of the ship power system and various working condition simulation models are used, combined with a small amount of real abnormal data and expert knowledge and experience, to generate normal and abnormal samples.

2. The method according to claim 1, characterized in that Multi-source and multi-modal state information fusion monitoring specifically includes: Based on the layout of the equipment and connecting pipelines in the ship's power system engine room, key areas that need to be monitored are analyzed. Infrared cameras, sound acquisition sensors, and vibration monitoring sensors are installed at appropriate locations on the system or equipment to monitor the status of the power system mechanical equipment and pipelines in real time. The sound and vibration signals are converted into images through signal conversion methods such as Mel spectrum conversion. The video and image information after sound and vibration conversion are processed through a large multimodal model. The video, sound, and vibration information are integrated to comprehensively judge whether the power system mechanical equipment and connecting pipelines have any abnormal conditions such as mechanical damage or deformation; Temperature, pressure, flow and other monitoring sensors are arranged at key locations on the power system pipelines to realize real-time monitoring of the system's thermal state. The temperature, pressure and flow under the system's operating state are integrated into image information that can be processed by a large multi-modal model to comprehensively judge whether the system's thermal state is normal.

3. The method according to claim 1, characterized in that The design of the small sample anomaly detection method includes: Generate normal samples of mechanical equipment's appearance and structure, as well as abnormal samples such as loosening and deformation of key screws; Generate normal and abnormal samples of rotating machinery vibration signals; Normal and abnormal sample generation of system operation status; The three methods mentioned above are used to generate normal and abnormal samples of the ship power system and fine-tune the CLIP model.

4. The method according to claim 3, characterized in that Normal and abnormal data generation for loosening and deformation of key screws in mechanical equipment specifically includes: Capture multi-angle images of equipment and pipelines in the area that needs to be monitored in the 3D model of the ship's power system, and add descriptions to the images; Based on existing images of pipeline deformation and loosening of key screws of mechanical equipment or expert knowledge, abnormal conditions of mechanical equipment such as pipeline deformation and loosening of key screws are simulated in a three-dimensional model, and images are captured as abnormal data samples.

5. The method according to claim 3, characterized in that The generation of normal and abnormal samples of rotating machinery vibration signals specifically includes: Vibration simulation analysis of rotating machinery is used to simulate and generate normal and abnormal vibration signals. Vibration frequency data is converted into Mel spectrum feature time-frequency graphs through Mel spectrum feature transformation, and the vibration frequency time series signal is converted into images. Based on existing rotating machinery vibration signals or expert knowledge, a Mel-spectrum feature time-frequency diagram of the abnormal equipment vibration frequency is generated and labeled "abnormal vibration of rotating machinery equipment", forming a series of image-text pairs.

6. The method according to claim 3, characterized in that The generation of normal and abnormal system operation status samples specifically includes: Convert temperature, pressure, and flow values ​​to 0-255 and convert them into images to suit CLIP model processing; The temperature, pressure and flow values ​​at various locations in the system are produced by the power system simulation model under normal and abnormal conditions; By integrating the above methods, "main power system heat dissipation is normal" is marked on the normal state value, and "main power system heat dissipation is abnormal" is marked on the abnormal state value.

7. The method according to claim 3, characterized in that Generate normal samples of the ship power system and fine-tune the CLIP model including: Convert the description of the anomaly into English and use it as the input of the CLIP model. Use the encoder to convert the image and text into the same space for comparison. The CLIP model optimizes the symmetric cross-entropy loss so that correct image and text pairs are closer in space and incorrect pairs are further away.

8. The method according to claim 1, characterized in that Also includes: The intelligent alarm method is designed. After processing the image through the CLIP model, the probability value of the similarity between the image and the abnormal description can be obtained. The abnormal description with the greatest similarity is selected as the abnormal information in the image. That is, the abnormal information in the image can be obtained through the CLIP model. Based on the abnormality description, the handling method can be preset through expert knowledge and experience, and the relevant personnel of the abnormality can be listed according to the ship power system operating procedures; Based on the list of relevant personnel of the exception, intelligent push is implemented to operators who need to participate in the processing or need to know the exception.

Images