Image-based inspection system for ferry cabins and anomaly localization method

Abstract

This invention discloses an image-based ferry cabin inspection system and anomaly localization method, belonging to the field of inspection and localization technology. The system includes: acquiring historical inspection logs of the ferry cabin; constructing a causal knowledge graph representing the causal relationships between various events within the cabin; real-time acquisition of inspection image sequences within the cabin; identifying the set of real-time events currently occurring within the cabin; calculating the probability of occurrence of the other events within a specified future time window; outputting anomalous events and their probabilities among future events with a probability exceeding a set threshold within the specified future time window as potential anomaly risk events; and locating the potential source location or associated region of the anomaly based on the output potential anomaly risk events and their corresponding causal paths. The advantages of this invention are: achieving comprehensive automated coverage of the core area of ​​the cabin, effectively reducing reliance on manual inspections, and comprehensively enhancing the proactive protection capabilities for ferry operation safety.

Description

Technical Field

[0001] This invention relates to the field of inspection and positioning technology, specifically to an image-based inspection system for ferry cabins and anomaly location methods. Background Technology

[0002] Currently, safety monitoring inside ferries mainly relies on fixed sensors, such as smoke detectors, temperature probes, and regular manual inspections. Sensors have limited monitoring range and can only trigger alarms based on preset physical quantity thresholds, failing to identify complex events. Manual inspections suffer from low timeliness, strong subjectivity, and insufficient coverage at night and in harsh environments. Both methods struggle to capture dynamically changing cabin conditions in real time. Especially in the event of emergencies such as fires or mechanical failures, existing systems can only respond passively after anomalies become apparent, unable to predict the evolution of the risk chain in advance.

[0003] In recent years, computer vision-based cabin monitoring systems have begun to be used, but they focus on single-point anomaly detection, such as identifying flames or people falling through images. These systems lack both a causal relationship model between historical and real-time events and the ability to proactively predict future risks. Due to the dense equipment and complex environment inside ferries, a single anomaly can trigger secondary disasters through equipment linkage and spatial proximity, such as localized leaks leading to electrical short circuits. Existing technologies, by neglecting the causal propagation mechanism between events, cannot pinpoint potential risk sources and make it difficult to upgrade from passive alarms to proactive protection. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides an image-based ferry cabin inspection system and anomaly localization method. This technical solution solves the problem that existing systems can only respond passively after anomalies become apparent and cannot predict the evolution of the risk chain in advance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A method for locating anomalies inside a ferry cabin based on image monitoring includes:

[0007] Obtain the historical inspection log inside the ferry cabin, which contains various historical events that occurred inside the cabin and their corresponding timestamps.

[0008] Based on the historical inspection logs, the causal correlation strength and / or chronological probability between different historical events are calculated to construct a causal knowledge graph that characterizes the causal correlation between various events in the cabin.

[0009] By deploying multiple image acquisition devices in key areas of the ferry cabin, real-time inspection image sequences are collected within the cabin.

[0010] The real-time acquired inspection image sequence is processed using an image recognition model to identify the set of real-time events currently occurring in the cabin. The set of real-time events includes possible event types and the comprehensive probability of each event type occurring.

[0011] Based on other events in the causal knowledge graph that are causally related to the current real-time event, and the corresponding causal relationship strength and / or temporal probability, calculate the probability of the occurrence of the other events within a specified future time window;

[0012] The abnormal times and probabilities of future events whose probability of occurrence exceeds a set threshold within a specified future time window will be output as potential events.

[0013] Determine whether the potential event is a defined abnormal event type. If so, mark the potential event as a potential abnormal risk event. If not, do not respond to the output of the potential event.

[0014] Based on the output of potential abnormal risk events and their corresponding causal paths, the potential source location or associated region of the anomaly can be located.

[0015] Preferably, the construction of the causal knowledge graph characterizing the causal relationships between various events within the cabin specifically includes:

[0016] Preprocess and define events in historical inspection logs;

[0017] Based on the timestamps of the events, statistical analysis is performed on the distribution and frequency of time intervals in which event A occurs before event B in the event sequence;

[0018] Based on the established time window and significance principle, calculate the causal weight of event A as a cause of event B;

[0019] Using events as nodes and causal relationships represented by causal weights that meet preset conditions as edges, a directed weighted graph is constructed to form the causal knowledge graph.

[0020] Preferably, the step of processing the real-time acquired inspection image sequence using an image recognition model to identify the set of real-time events currently occurring inside the cabin, wherein the set of real-time events includes possible event types and the comprehensive probability of each event type occurring, specifically including:

[0021] Based on the trained image recognition model, event recognition is performed on each inspection image in the inspection image sequence to obtain the possible event types and the probability of each event type occurring for each inspection image.

[0022] Based on the recognition results of all inspection images in the inspection image sequence, determine the comprehensive probability of each possible event type.

[0023] By summarizing all possible event types and the combined probability of each event type occurring, a real-time event set is constructed.

[0024] Specifically, the calculation process is as follows:

[0025]

[0026] In the formula, Let X be the probability of the Xth possible event type occurring. For inspection image sequences, For the e-th element in the inspection image sequence, Let X be the probability of the Xth possible event type corresponding to the eth element in the inspection image sequence.

[0027] Preferably, the step of calculating the probability of occurrence of other events within a specified future time window based on other events in the causal knowledge graph that are causally related to the current real-time event, and the corresponding causal association strength and / or temporal probability, specifically includes:

[0028] Based on the current real-time event set and causal knowledge graph, event chain reasoning is performed to calculate the probability of occurrence of subsequent node events on the event chain within the future time window. The probability of occurrence is the sum of the causal weights of each link in the path.

[0029] The specific calculation formula is as follows:

[0030]

[0031] In the formula, Let Y be the probability of the Yth subsequent node event occurring within the future time window. Let L be the probability of the Lth possible event type occurring. Let Y be the set of event chain sources corresponding to the Yth subsequent node event. For the event chain from event L to event Y, For adjacent node events, for Causal weights between them.

[0032] Preferably, locating the potential source location or associated region of the anomaly based on the output potential abnormal risk events and their corresponding causal paths specifically includes:

[0033] Based on the source events of all event chains corresponding to potential abnormal risk events, analyze the potential risk contribution value of each real-time event to the potential abnormal risk events;

[0034] Determine the associated region for each real-time event;

[0035] The potential risk value of the associated region is obtained by summing the potential risk contributions of all real-time events corresponding to the associated region to potential abnormal risk events.

[0036] Output the associated regions in descending order.

[0037] Furthermore, an image-based ferry cabin inspection system is proposed to realize the image-based ferry cabin anomaly localization method described above, including:

[0038] The image acquisition unit is deployed in key areas inside the ferry cabin to collect real-time images of the cabin inspection sequence.

[0039] The historical log storage unit stores the ferry cabin inspection logs containing historical event types and their timestamps.

[0040] The causal graph construction unit, connected to the historical log storage unit, is used to calculate the causal correlation strength between events based on the temporal relationship of historical events, and generate a weighted causal knowledge graph with events as nodes and causal relationships as edges.

[0041] A real-time event recognition unit, connected to the image acquisition unit, is configured with an image recognition model for recognizing the current event type in an image sequence and outputting a set of real-time events with probability weights.

[0042] The event prediction unit, which connects the causal graph construction unit and the real-time event identification unit, is used to deduce the probability of other events occurring within a specified time window in the future based on the association path between the real-time event set and the causal knowledge graph.

[0043] An anomaly determination unit, connected to the event prediction unit, is used to filter future events with a probability of occurrence exceeding a preset threshold and to mark potential anomaly risk events based on a predefined anomaly type library;

[0044] An anomaly localization unit, connected to the anomaly determination unit, is used to trace the related areas and output a risk area sorting list based on the causal relationship path of potential anomaly risk events.

[0045] The alarm output unit is used to publish alarm information that includes potential abnormal risk events and risk areas.

[0046] Optionally, the causal mapping unit includes:

[0047] The event preprocessing subunit is used to categorize event types in the historical logs;

[0048] The causal analysis subunit is used to count the sequential frequency of events and calculate causal weights.

[0049] Optionally, the real-time event recognition unit includes:

[0050] A separate identification subunit is used to perform individual event identification on each inspection image in the inspection image sequence to obtain the possible event types and the probability of each event type occurring for each inspection image.

[0051] The event probability fusion subunit is used to generate a comprehensive probability value for the event type by combining the recognition results of multiple frames of images.

[0052] Optionally, the anomaly localization unit includes:

[0053] The source contribution calculation subunit is used to quantify the contribution of the source event in the causal chain to the risk;

[0054] The area mapping sub-unit associates events with pre-defined physical areas within the ship's cabin;

[0055] The regional risk ranking sub-unit aggregates risk contribution values ​​by region and outputs them in descending order.

[0056] Optionally, the image acquisition unit includes:

[0057] A multi-view camera array with spatiotemporal markers is deployed on the roof and at equipment nodes;

[0058] The image synchronization component adds a location tag and timestamp to each frame of the image.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] This invention revolutionizes the safety monitoring mechanism inside ferries by constructing a causal knowledge graph of events within the cabin and integrating real-time image sequence analysis. It not only identifies various dynamic anomalies in real time but also predicts the evolution of potential risk chains through a causal reasoning model between historical and real-time events. This allows the system to proactively warn of secondary disasters before they manifest, significantly advancing the risk mitigation window. Simultaneously, based on the reverse tracing capability of the event causal chain, it accurately locates the physical source of potential hazards, completely solving the blind spot problem of traditional monitoring that makes it difficult to trace the root cause of anomalies, and significantly improving accident investigation efficiency. Furthermore, the collaborative perception of multi-dimensional states such as equipment malfunctions, environmental changes, and personnel behavioral risks achieves comprehensive automated coverage of the core areas of the cabin, effectively reducing reliance on manual inspections and comprehensively enhancing the proactive protection capabilities for ferry operation safety. Attached Figure Description

[0061] Figure 1 The flowchart is a method for locating anomalies inside a ferry cabin based on image monitoring, as proposed in this invention.

[0062] Figure 2This is a flowchart of the method for constructing a causal knowledge graph proposed in this invention;

[0063] Figure 3 This is a flowchart of the method for identifying the set of real-time events currently occurring inside the cabin, as proposed in this invention.

[0064] Figure 4 This is a flowchart of the method for locating the potential source location or associated region of an anomaly proposed in this invention. Detailed Implementation

[0065] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0066] Reference Figure 1 As shown, a method for locating anomalies inside a ferry cabin based on image monitoring includes:

[0067] Obtain the historical inspection log inside the ferry cabin. The historical inspection log contains various historical events that occurred inside the cabin and their corresponding timestamps.

[0068] Based on historical event data accumulated over a long period of time, we can deeply explore the operational patterns inside the cabin, providing a reliable data foundation for building causal relationship models and significantly improving the accuracy and reliability of subsequent risk predictions.

[0069] Based on historical inspection logs, calculate the causal relationship strength and / or chronological probability between different historical events, and construct a causal knowledge graph that represents the causal relationship between various events in the cabin.

[0070] By quantifying the causal relationships between events, discrete historical events are transformed into a structured knowledge network, breaking through the limitations of traditional monitoring methods that rely solely on single event responses, and laying the core logical framework for the accurate prediction of real-time risk evolution.

[0071] By deploying multiple image acquisition devices in key areas of the ferry cabin, real-time inspection image sequences are collected within the cabin.

[0072] Multi-view, high-density coverage of the core area, real-time capture of dynamic environmental changes, overcoming sensor monitoring blind spots and the lag of manual inspection, and achieving continuous perception of the entire cabin's operational status.

[0073] The image recognition model is used to process the real-time collected inspection image sequence to identify the set of real-time events currently occurring in the cabin. The set of real-time events includes possible event types and the comprehensive probability of event types occurring.

[0074] By fusing recognition results from multiple frames of images to generate probabilistic event conclusions, the false alarm rate is significantly reduced, such as distinguishing between normal water mist and leaking steam, while also enhancing the robustness of identification in complex scenes.

[0075] Based on other events in the causal knowledge graph that are causally related to the current real-time event, and the corresponding causal relationship strength and / or temporal probability, calculate the probability of other events occurring within a specified future time window;

[0076] By combining real-time dynamics with historical causal chains, the system proactively predicts the evolution path of secondary disasters, such as leakage → short circuit → fire, achieving a fundamental shift from post-event warnings to pre-event predictions and significantly extending the risk mitigation window.

[0077] The abnormal times and probabilities of future events whose probability of occurrence exceeds a set threshold within a specified future time window will be output as potential events.

[0078] Determine whether the potential event is a defined abnormal event type. If so, mark the potential event as a potential abnormal risk event. If not, do not respond to the output of the potential event.

[0079] Filtering low-risk events based on a pre-defined anomaly type library, focusing on real high-risk targets such as electrical fires and structural collapses, avoiding interference from invalid alarms, and improving the efficiency and reliability of system decision-making;

[0080] Based on the output of potential abnormal risk events and their corresponding causal paths, the potential source location or associated region of the anomaly can be located.

[0081] By tracing the risk transmission chain back to its physical source, such as locating leaks or faulty equipment, this system completely solves the problem of traditional systems failing to identify the source of the fire, guiding precise and rapid response and minimizing accident losses.

[0082] Reference Figure 2 As shown, constructing a causal knowledge graph representing the causal relationships between various events within the cabin specifically includes:

[0083] Preprocess and define events in historical inspection logs;

[0084] Based on the timestamps of the events, statistical analysis is performed on the distribution and frequency of time intervals in which event A occurs before event B in the event sequence;

[0085] Based on the set time window and significance principle, the causal weight of event A as the cause of event B is calculated. Specifically, in some preferred embodiments, the causal weight of event A as the cause of event B can be obtained by comparing the number of times event B occurs within the time window after event A occurs with the sum of the number of times all events occur within the time window after event A occurs.

[0086] Using events as nodes and causal relationships represented by causal weights that meet preset conditions as edges, a directed weighted graph is constructed to form a causal knowledge graph.

[0087] By constructing a causal knowledge graph through quantifying the temporal dependencies between historical events, three core breakthroughs are achieved: First, the proportion of causal associations within an event window is used as the weighting criterion, such as the percentage of the frequency of event B occurring within a specific time period after event A. This objectively isolates accidental temporal relationships from a statistical perspective, significantly improving the interpretability and accuracy of causal determination and avoiding the bias of traditional expert systems that rely on subjective experience rules. Second, the constructed dynamic directed weighted graph can adaptively integrate newly added historical data, enabling the knowledge graph to continuously evolve and optimize the modeling accuracy of risk propagation patterns in complex systems within the cabin. Finally, this graph provides a physically and logically complete reasoning framework for risk extrapolation of subsequent real-time events, ensuring that the prediction results of key causal chains such as "equipment leakage" to "electrical short circuit" have clear engineering traceability. This fundamentally solves the trust crisis in operation and maintenance decisions caused by black-box prediction models and lays the technical foundation for the engineering implementation of proactive protection systems.

[0088] Reference Figure 3 As shown, an image recognition model is used to process the real-time acquired inspection image sequence to identify the set of real-time events currently occurring inside the cabin. The set of real-time events includes possible event types and the comprehensive probability of each event type occurring, specifically including:

[0089] Based on the trained image recognition model, event recognition is performed on each inspection image in the inspection image sequence to obtain the possible event types and the probability of each event type occurring for each inspection image. The image recognition model can be trained using a convolutional neural network. The softmax layer converts the probability of the event recognized by the image recognition model into a probability value between 0 and 1. Convolutional neural networks are a mature technology in this field. The training process of the image recognition model does not involve the core improvement points of this solution, so it will not be described in detail here.

[0090] Based on the recognition results of all inspection images in the inspection image sequence, determine the comprehensive probability of each possible event type.

[0091] By summarizing all possible event types and the combined probability of each event type occurring, a real-time event set is constructed.

[0092] Specifically, the calculation process is as follows:

[0093]

[0094] In the formula, Let X be the probability of the Xth possible event type occurring. For inspection image sequences, For the e-th element in the inspection image sequence, Let X be the probability of the Xth possible event type corresponding to the eth element in the inspection image sequence.

[0095] By employing a probabilistic fusion mechanism based on multi-frame images, the robustness and confidence of event recognition in complex cabin environments are significantly improved. First, comprehensive probability calculations are performed based on the recognition results of sequential images. For example, by capturing continuous state changes of equipment leakage from multiple angles, misjudgments caused by viewpoint obstruction, light interference, or instantaneous actions in single-frame images are effectively eliminated, raising the confidence of critical event judgments to an operational level. Second, this probabilistic fusion model can adapt to complex interference environments such as dynamic vibrations, sudden changes in light, and equipment obstruction during ferry operation, ensuring the output of highly reliable event sets in mobile scenarios. Finally, the generated probabilistic event descriptions provide quantitative input for subsequent causal chain deduction, making "78% probability of event occurrence" more accurate than the binary judgment of "event occurrence" in supporting risk classification decisions. This fundamentally solves the false alarm problem of image monitoring technology in industrial applications, providing highly reliable original decision-making basis for predictive protection systems.

[0096] Based on other events in the causal knowledge graph that are causally related to the current real-time event, and their corresponding causal relationship strength and / or temporal probability, the probability of these other events occurring within a specified future time window is calculated, specifically including:

[0097] Based on the current real-time event set and causal knowledge graph, event chain reasoning is performed to calculate the probability of subsequent node events on the event chain occurring within a future time window. The probability of occurrence is the sum of the causal weights of each link in the path.

[0098] The specific calculation formula is as follows:

[0099]

[0100] In the formula, Let Y be the probability of the Yth subsequent node event occurring within a future time window. Let L be the probability of the Lth possible event type occurring. Let Y be the set of event chain sources corresponding to the Yth subsequent node event. For the event chain from event L to event Y, For adjacent node events, for Causal weights between them.

[0101] Through the aggregation mechanism of causal weights in the event chain, a quantitative prediction of complex cascading risks has been achieved for the first time in the field of ship monitoring. Based on the dynamic coupling of real-time event probabilities and historical causal weights, the evolution path of multi-level secondary disasters is accurately captured, such as the transmission chain of equipment leakage → electrical short circuit → cabin smoke, breaking through the technical bottleneck of traditional models that can only judge direct causal relationships. At the same time, the mathematical essence of probability product enhances the sensitivity to long-cycle, weakly correlated risks. For example, the probability of an initial weak leakage triggering a fire after three levels of transmission jumps, enabling the system to warn of systemic risks that traditional methods cannot identify in the early stages of potential hazards. Finally, the output event chain probability value provides a quantitative basis for operation and maintenance decisions, such as prioritizing the handling of "leakage-short circuit" combination events with weighted probabilities exceeding the threshold, fundamentally avoiding the misallocation of protection resources caused by misjudgment of discrete events, and significantly improving the accuracy of proactive prevention and control in high-risk scenarios.

[0102] Reference Figure 4 As shown, based on the output of potential abnormal risk events and their corresponding causal paths, locating the potential source location or associated region of the anomaly specifically includes:

[0103] Based on the source events of all event chains corresponding to potential abnormal risk events, analyze the potential risk contribution value of each real-time event to the potential abnormal risk events;

[0104] Determine the associated region for each real-time event;

[0105] The potential risk value of the associated region is obtained by summing the potential risk contributions of all real-time events corresponding to the associated region to potential abnormal risk events.

[0106] Output the associated regions in descending order.

[0107] Specifically, the potential risk contribution value of each real-time event to a potential abnormal risk event is... The potential risk contribution value reflects the risk probability orientation value of real-time events, which originate from real-time events, ultimately leading to potential abnormal risk events. The larger the value, the greater the causal orientation relationship between real-time events and the occurrence of potential abnormal risk events.

[0108] By employing a regional accumulation and sorting mechanism for risk contribution values, a leapfrog breakthrough is achieved, moving from abstract event early warning to precise physical spatial positioning. First, based on causal chain reverse tracing, the contribution intensity of each source event to the terminal risk is calculated, such as the weighted guiding value of oil pump leakage to electrical fire. Secondary related events are quantitatively separated to focus on core risk sources, such as accurately distinguishing between leakage sources causing short circuits and unrelated pipe condensate. Second, by mapping regional risk value accumulation to physical space, the industry has achieved a cognitive upgrade from "where the anomaly occurs" to "where the source of the hidden danger is" for the first time, completely solving the pain point that traditional monitoring and alarms still require several hours of manual investigation. Finally, regional sequences are output in descending order of risk values, directly guiding the efficient deployment of maintenance resources, improving maintenance response efficiency, and transforming the timeliness of handling accident hazards from "post-event rescue" to "source prevention."

[0109] Furthermore, based on the aforementioned image-based anomaly localization method for ferry cabins, this solution also proposes an image-based ferry cabin inspection system, including:

[0110] The image acquisition unit is deployed in key areas inside the ferry cabin to collect real-time images of the cabin inspection sequence.

[0111] The image acquisition unit includes a multi-view camera group with spatiotemporal markers, deployed on the top of the cabin and equipment nodes, as well as an image synchronization component that adds location tags and timestamps to each frame of image;

[0112] The historical log storage unit stores the ferry cabin inspection logs containing historical event types and their timestamps.

[0113] The causal graph construction unit, connected to the historical log storage unit, is used to calculate the strength of causal associations between events based on the temporal relationship of historical events, and generate a weighted causal knowledge graph with events as nodes and causal relationships as edges;

[0114] The causal graph construction unit includes:

[0115] The event preprocessing subunit is used to categorize event types in the historical logs;

[0116] The causal analysis subunit is used to statistically analyze the temporal frequency of events and calculate causal weights.

[0117] The real-time event recognition unit is connected to the image acquisition unit and is equipped with an image recognition model to identify the current event type in the image sequence and output a set of real-time events with probability weights.

[0118] Real-time event recognition unit:

[0119] The individual identification subunit is used to perform individual event identification on each inspection image in the inspection image sequence to obtain the possible event types and the probability of each event type occurring for each inspection image.

[0120] The event probability fusion subunit is used to generate a comprehensive probability value for the event type by combining the recognition results of multiple frames of images;

[0121] The event prediction unit connects the causal graph construction unit and the real-time event identification unit. It is used to deduce the probability of other events occurring within a specified time window in the future based on the association path between the real-time event set and the causal knowledge graph.

[0122] The anomaly determination unit, connected to the event prediction unit, is used to filter future events whose probability of occurrence exceeds a preset threshold and to mark potential anomaly risk events based on a predefined anomaly type library.

[0123] The anomaly localization unit, connected to the anomaly determination unit, is used to trace the related areas based on the causal relationship path of potential anomaly risk events and output a sorted list of risk areas.

[0124] The anomaly localization unit includes:

[0125] The source contribution calculation subunit is used to quantify the contribution of the source event in the causal chain to the risk;

[0126] The area mapping sub-unit associates events with pre-defined physical areas within the ship's cabin;

[0127] The regional risk ranking sub-unit aggregates risk contribution values ​​by region and outputs them in descending order.

[0128] The alarm output unit is used to publish alarm information that includes potential abnormal risk events and risk areas.

[0129] In summary, the advantages of this invention are as follows: By constructing a causal knowledge graph of events within the cabin and integrating real-time image sequence analysis, it revolutionizes the safety monitoring mechanism within the ferry cabin. It not only identifies various dynamic abnormal states in real time but also predicts the evolution trend of potential risk chains through a causal reasoning model between historical and real-time events. This allows the system to proactively issue warnings before secondary disasters manifest, significantly advancing the risk mitigation window. Simultaneously, based on the reverse tracing capability of the event causal chain, it accurately locates the physical source of potential hazards, completely solving the blind spot problem of tracing the root cause of anomalies in traditional monitoring and significantly improving accident investigation efficiency. Furthermore, the collaborative perception of multi-dimensional states such as equipment malfunctions, environmental changes, and personnel behavioral risks achieves comprehensive automated coverage of the core cabin area, effectively reducing reliance on manual inspections and comprehensively enhancing the proactive protection capability for ferry operation safety.

[0130] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for locating anomalies inside a ferry cabin based on image monitoring, characterized in that, include: Obtain the historical inspection log inside the ferry cabin, which contains various historical events that occurred inside the cabin and their corresponding timestamps. Based on the historical inspection logs, the causal correlation strength and / or chronological probability between different historical events are calculated to construct a causal knowledge graph that characterizes the causal correlation between various events in the cabin. By deploying multiple image acquisition devices in key areas of the ferry cabin, real-time inspection image sequences are collected within the cabin. The real-time acquired inspection image sequence is processed using an image recognition model to identify the set of real-time events currently occurring in the cabin. The set of real-time events includes possible event types and the comprehensive probability of each event type occurring. Based on other events in the causal knowledge graph that are causally related to the current real-time event, and the corresponding causal relationship strength and / or temporal probability, calculate the probability of the occurrence of the other events within a specified future time window; The abnormal times and probabilities of future events whose probability of occurrence exceeds a set threshold within a specified future time window will be output as potential events. Determine whether the potential event is a defined abnormal event type. If so, mark the potential event as a potential abnormal risk event. If not, do not respond to the output of the potential event. Based on the output of potential abnormal risk events and their corresponding causal paths, locate the potential source location or associated region of the anomaly. The process of using an image recognition model to process the real-time acquired inspection image sequence identifies a set of real-time events currently occurring inside the cabin. This set of real-time events includes possible event types and the combined probability of each event type occurring. Specifically, it includes: Based on the trained image recognition model, event recognition is performed on each inspection image in the inspection image sequence to obtain the possible event types and the probability of each event type occurring for each inspection image. Based on the recognition results of all inspection images in the inspection image sequence, determine the comprehensive probability of each possible event type. By summarizing all possible event types and the combined probability of each event type occurring, a real-time event set is constructed. Specifically, the calculation process is as follows: ； In the formula, Let X be the probability of the Xth possible event type occurring. For inspection image sequences, For the e-th element in the inspection image sequence, Let X be the probability of the possible event type corresponding to the e-th element in the inspection image sequence. The step of calculating the probability of occurrence of other events within a specified future time window based on other events in the causal knowledge graph that are causally related to the current real-time event, and the corresponding causal relationship strength and / or temporal probability, specifically includes: Based on the current real-time event set and causal knowledge graph, event chain reasoning is performed to calculate the probability of occurrence of subsequent node events on the event chain within the future time window. The probability of occurrence is the sum of the causal weights of each link on the path. The specific calculation formula is as follows: ； In the formula, Let Y be the probability of the Yth subsequent node event occurring within the future time window. Let L be the probability of the Lth possible event type occurring. Let Y be the set of event chain sources corresponding to the Yth subsequent node event. For the event chain from event L to event Y, For adjacent node events in the middle, for Causal weights between them.

2. The method for anomaly localization inside a ferry cabin based on image monitoring according to claim 1, characterized in that, The construction of a causal knowledge graph representing the causal relationships between various events within the cabin specifically includes: Preprocess and define events in historical inspection logs; Based on the timestamps of the events, statistical analysis is performed on the distribution and frequency of time intervals in which event A occurs before event B in the event sequence; Based on the established time window and significance principle, calculate the causal weight of event A as a cause of event B; Using events as nodes and causal relationships represented by causal weights that meet preset conditions as edges, a directed weighted graph is constructed to form the causal knowledge graph.

3. The method for anomaly localization inside a ferry cabin based on image monitoring according to claim 2, characterized in that, The process of locating the potential source location or associated region of an anomaly based on the output of potential abnormal risk events and their corresponding causal paths specifically includes: Based on the source events of all event chains corresponding to potential abnormal risk events, analyze the potential risk contribution value of each real-time event to the potential abnormal risk events; Determine the associated region for each real-time event; The potential risk value of the associated region is obtained by summing the potential risk contributions of all real-time events corresponding to the associated region to potential abnormal risk events. Output the associated regions in descending order.

4. A ferry cabin inspection system based on image monitoring, characterized in that, The method for locating anomalies inside a ferry cabin based on image monitoring as described in any one of claims 1-3 includes: The image acquisition unit is deployed in key areas inside the ferry cabin to collect real-time images of the cabin inspection sequence. The historical log storage unit stores the ferry cabin inspection logs containing historical event types and their timestamps. The causal graph construction unit, connected to the historical log storage unit, is used to calculate the causal correlation strength between events based on the temporal relationship of historical events, and generate a weighted causal knowledge graph with events as nodes and causal relationships as edges. A real-time event recognition unit, connected to the image acquisition unit, is configured with an image recognition model for recognizing the current event type in an image sequence and outputting a set of real-time events with probability weights. The event prediction unit, which connects the causal graph construction unit and the real-time event identification unit, is used to deduce the probability of other events occurring within a specified time window in the future based on the association path between the real-time event set and the causal knowledge graph. An anomaly determination unit, connected to the event prediction unit, is used to filter future events with a probability of occurrence exceeding a preset threshold and to mark potential anomaly risk events based on a predefined anomaly type library; An anomaly localization unit, connected to the anomaly determination unit, is used to trace the related areas and output a risk area sorting list based on the causal relationship path of potential anomaly risk events. The alarm output unit is used to publish alarm information that includes potential abnormal risk events and risk areas.

5. The ferry cabin inspection system based on image monitoring according to claim 4, characterized in that, The causal mapping construction unit includes: The event preprocessing subunit is used to categorize event types in the historical logs; The causal analysis subunit is used to count the sequential frequency of events and calculate causal weights.

6. The ferry cabin inspection system based on image monitoring according to claim 4, characterized in that, The real-time event recognition unit includes: A separate identification subunit is used to perform individual event identification on each inspection image in the inspection image sequence to obtain the possible event types and the probability of each event type occurring for each inspection image. The event probability fusion subunit is used to generate a comprehensive probability value for the event type by combining the recognition results of multiple frames of images.

7. The ferry cabin inspection system based on image monitoring according to claim 4, characterized in that, The anomaly localization unit includes: The source contribution calculation subunit is used to quantify the contribution of the source event in the causal chain to the risk; The area mapping sub-unit associates events with pre-defined physical areas within the ship's cabin; The regional risk ranking sub-unit aggregates risk contribution values ​​by region and outputs them in descending order.

8. The ferry cabin inspection system based on image monitoring according to claim 4, characterized in that, The image acquisition unit includes: A multi-view camera array with spatiotemporal markers is deployed on the roof and at equipment nodes; The image synchronization component adds a location tag and timestamp to each frame of the image.

Images