A method and device for early warning of the risk of falling into the water for an offshore wind power maintenance personnel

Abstract

The application provides a method and device for early warning of falling risk of offshore wind power operation and maintenance personnel, which is used for improving the explainability of falling risk early warning and reducing the false positive rate. The method comprises the following steps: forming multi-source fusion falling risk data; updating the attributes of key entities according to the multi-source fusion falling risk data, and converting the corresponding attributes into real-time falling risk facts with time identifiers when the updated attributes meet the preset risk state triggering conditions; establishing the time sequence correlation relationship between the real-time falling risk facts according to the occurrence time and spatial position of each real-time falling risk fact; fusing the key entities, static correlation relationship, real-time falling risk facts and time information to construct a time sequence falling risk knowledge graph; reasoning the falling risk evolution trend under the joint action of multiple factors according to the historical evolution law in the time sequence falling risk knowledge graph and the real-time falling risk fact, generating explainable falling risk early warning information and outputting.

Description

Technical Field

[0001] This invention belongs to the field of detecting falls into the water by offshore wind power operation and maintenance personnel, and specifically relates to a method and device for early warning of the risk of falls into the water by offshore wind power operation and maintenance personnel. Background Technology

[0002] With the large-scale development of offshore wind power and its gradual expansion into deep-sea areas, the offshore wind power operation and maintenance environment is characterized by its distance from land, complex environmental conditions, and limited working space. When operating in this complex environment, maintenance personnel are at risk of falling overboard due to a combination of factors, including their condition, location, environmental conditions, and the specific tasks at hand. Once overboard, the unique geographical conditions at sea make detection, location, and rescue extremely difficult, potentially leading to serious safety accidents. Therefore, real-time perception and accurate early warning of the risk of maintenance personnel falling overboard are crucial to reducing the probability of such accidents and ensuring the safety of offshore wind power operation and maintenance.

[0003] In recent years, the rapid development of wearable sensing technology and artificial intelligence technology has provided technical support for the above-mentioned needs. Existing methods typically collect multi-source data such as vital signs and behavioral states of personnel, and use machine learning models to establish a mapping relationship between risk states and multi-source data to achieve the identification and prediction of the risk of falling into water. However, this type of purely data-driven method has two shortcomings: First, the decision-making process within the model is not transparent, making it difficult to explicitly describe the complex correlation between personnel status, spatial location, environmental factors, and work tasks, resulting in a lack of interpretability in the warning results; second, when multiple risk factors act together, this method is difficult to effectively characterize the temporal correlation and risk evolution process between different factors, and is prone to false alarms triggered by the anomaly of a single indicator, or missed alarms due to the coupling of multiple factors. The root cause of these shortcomings is that existing methods lack a structured knowledge carrier that can uniformly express the complex correlation between personnel, space, environment, task, and risk.

[0004] Therefore, how to construct an explicit expression of the above-mentioned relationships and realize an interpretable early warning of the risk of falling into the water based on its temporal evolution characteristics has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a method and device for early warning of the risk of falling overboard for offshore wind power operation and maintenance personnel, which is used to improve the interpretability of the risk warning and reduce the false alarm rate.

[0006] A method for early warning of the risk of offshore wind power operation and maintenance personnel falling into the water includes:

[0007] Based on the status information of maintenance personnel, spatial status information, and environmental and task status information, multi-source fusion data on water risk is generated.

[0008] Obtain a pre-constructed initial knowledge representation of the risk of falling into water. The initial knowledge representation of the risk of falling into water is pre-constructed based on historical risk event data and domain prior knowledge, and includes multiple key entities that affect the risk of falling into water and the static correlation between each key entity.

[0009] Based on the multi-source fusion water risk data, the attributes of the key entities are updated. When the updated attributes meet the preset risk state triggering conditions, the corresponding attributes are converted into real-time water risk facts with time stamps.

[0010] Based on the occurrence time and spatial location of each real-time water-fall risk fact, establish the temporal correlation between each real-time water-fall risk fact;

[0011] By integrating the key entities, the static relationships, the real-time facts of water-falling risk and their time information, a time-series water-falling risk knowledge graph is constructed.

[0012] Based on the historical evolution patterns in the time-series knowledge graph of falling water risk and the real-time facts of falling water risk, the evolution trend of falling water risk under the combined effect of multiple factors is inferred, and explainable falling water risk early warning information is generated and output. The explainable falling water risk early warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status association paths.

[0013] Preferably, the step of updating the attributes of the key entity based on the multi-source fused fall-in-water risk data, and converting the corresponding attributes into real-time fall-in-water risk facts with time stamps when the updated attributes meet the preset risk state triggering conditions, includes:

[0014] Based on the status information of maintenance personnel in the multi-source fused water risk data, the numerical values ​​of the personnel status attribute set associated with the personnel entity are updated; the personnel status attribute set includes vital sign status attributes and behavioral status attributes, and the time identifier of this update is recorded.

[0015] Based on the spatial state information in the multi-source fusion water-fall risk data, the spatial risk attribute set associated with the spatial entities is numerically updated, and the time identifier of this update is recorded.

[0016] Based on the environmental and task status information in the multi-source fusion water risk data, the environmental status attribute set associated with the environmental entity and the task attribute set associated with the task entity are updated numerically, and the time identifier of this update is recorded.

[0017] According to the preset risk status determination rules, the updated attributes of each key entity are analyzed. When the attribute of any key entity or the relationship between entities meets the preset risk status triggering conditions, the corresponding attribute or relationship is converted into a real-time drowning risk fact with a time stamp.

[0018] Preferably, the step of establishing the temporal correlation between the real-time fall-in-water risk facts based on their occurrence time and spatial location includes:

[0019] Obtain the entity type, relationship type, occurrence time, and spatial location corresponding to each of the aforementioned real-time water-fall risk facts;

[0020] Based on the occurrence time t of any two of the aforementioned real-time risk of falling into the water facts i and t j The time correlation coefficient is calculated using the time decay function. :

[0021]

[0022] Where k is a preset time decay coefficient, and the time correlation coefficient is... Used to characterize the time dependency in the continuous evolution of risk states;

[0023] Based on the spatial location p corresponding to any two real-time water risk facts i and p j Calculate the spatial correlation coefficient :

[0024]

[0025] The spatial correlation coefficient Used to characterize the spatial constraint relationship between changes in personnel activity areas and hazardous areas;

[0026] Based on the time correlation coefficient and the spatial correlation coefficient Calculate the overall correlation strength between any two of the aforementioned real-time fall-in-water risk facts. :

[0027] ;

[0028] in, The first preset weighting coefficient, The second preset weighting coefficient, This is the offset of the overall association strength for the preset temporal association relationships;

[0029] When the comprehensive correlation strength When the value exceeds a preset correlation threshold, a temporal correlation is established between the two real-time drowning risk facts, and the overall correlation strength is... As the association weight of the corresponding temporal relationship.

[0030] Preferably, the step of reasoning about the evolution trend of falling water risk under the combined effect of multiple factors, based on the historical evolution patterns in the time-series falling water risk knowledge graph and the real-time falling water risk facts, and generating and outputting interpretable falling water risk early warning information includes:

[0031] Based on the real-time risk facts of falling into the water at the current moment, the corresponding risk evolution pattern is matched in the time-series risk knowledge graph to determine the evolution trend of the current risk status of the maintenance personnel.

[0032] Based on the evolution trend, the temporal correlation between each key entity in the temporal knowledge graph of drowning risk is obtained, information is disseminated along the temporal correlation path for different risk factors, comprehensive risk correlation characteristics are obtained, and the correlation strength of the current risk state to the target risk state is obtained based on the comprehensive risk correlation characteristics.

[0033] Based on the correlation strength, the current risk status characteristics are fused with the comprehensive risk correlation characteristics and input into the drowning risk prediction model to calculate the probability that maintenance personnel will enter the target drowning risk state within a preset future time range.

[0034] Based on the contribution of each key entity to the probability, the key entities affecting the change of the current risk state are determined, the corresponding risk state association paths are extracted, and interpretable drowning risk warning information is generated and output.

[0035] Preferably, the step of determining the evolution trend of the current operation and maintenance personnel's risk status by matching the corresponding risk evolution pattern in the time-series water-fall risk knowledge graph based on the real-time water-fall risk facts at the current moment includes:

[0036] Based on the real-time risk of falling into the water at the current moment, obtain the corresponding risk entity type, entity relationship type, time of occurrence, and risk association weight;

[0037] Using the risk entity type and the entity association type as search criteria, historical risk facts with the same risk entity type and the same entity association type as the current risk status are extracted from the time-series drowning risk knowledge graph and arranged in chronological order to form a historical risk fact sequence.

[0038] Calculate the similarity between the temporal features corresponding to the current risk state and the temporal features corresponding to each candidate risk evolution pattern in the historical risk fact sequence;

[0039] The candidate risk evolution pattern with the highest similarity is identified as the matching risk evolution pattern, and the evolution trend corresponding to the current risk status of the operation and maintenance personnel is determined based on the matching risk evolution pattern.

[0040] Preferably, the steps of obtaining the temporal correlation relationships between the key entities in the temporal knowledge graph of drowning risk according to the evolution trend, propagating information about different risk factors along the temporal correlation path to obtain comprehensive risk correlation characteristics, and obtaining the correlation strength of the evolution from the current risk state to the target risk state based on the comprehensive risk correlation characteristics include:

[0041] Based on the aforementioned evolutionary trend, the key entities involved in the risk evolution are identified, including personnel entities, space entities, environmental entities, mission entities, and risk status entities.

[0042] Obtain the temporal association relationships and corresponding temporal association paths among the key entities involved in the risk evolution in the temporal knowledge graph of drowning risk.

[0043] Along the time-series association path, the risk factor characteristics of each key entity are sequentially transmitted to adjacent entities. When any entity receives risk factor characteristics transmitted by multiple adjacent entities, they are weighted and summed according to the risk association weights corresponding to each association relationship to form the comprehensive risk association characteristics of that entity.

[0044] Based on the comprehensive risk correlation characteristics, the trend of risk state change under the combined effect of different risk factors is analyzed to obtain the correlation strength of the evolution of the current risk state to the target risk state.

[0045] Preferably, the steps of determining the key entities affecting the change of the current risk state based on the contribution of each key entity to the probability, extracting the corresponding risk state association paths, generating and outputting interpretable drowning risk warning information include:

[0046] The contribution attribution method is used to calculate the degree of contribution of each key entity in the time-series knowledge graph of drowning risk to the probability.

[0047] The key entities are ranked according to their contribution level, and those whose contribution level exceeds a preset threshold are identified as key entities that affect the change of the current risk status.

[0048] Based on the key entities, corresponding risk facts, and relationships between entities, the risk state association path from the initial risk state to the current risk state is extracted from the time-series drowning risk knowledge graph.

[0049] Generate and output interpretable fall-in-water risk warning information, which includes risk level, risk factors and their contribution ranking, risk facts and risk status association path.

[0050] Preferably, the method further includes:

[0051] Based on the deviation between the output of the explainable water risk warning information and the actual risk status change results, the association weights between the key entities and the water risk status entities in the time-series water risk knowledge graph are updated.

[0052] Based on the updated relationships and newly added real-time water risk facts, the key entities, relationships, and risk evolution patterns in the time-series water risk knowledge graph are updated to optimize the time-series water risk knowledge graph.

[0053] Preferably, the key entities include personnel entities, space entities, environmental entities, mission entities, and water risk status entities;

[0054] The personnel entity is used to represent personnel objects participating in offshore wind power operation and maintenance activities, and is associated with a personnel status attribute set, which includes vital sign status attributes and behavioral status attributes.

[0055] The spatial entity is used to characterize the physical space and its spatial risk characteristics involved in offshore wind power operation and maintenance, and to associate a set of spatial risk attributes, which includes a hazard level.

[0056] The environmental entities are used to characterize external environmental factors that affect the safety of offshore wind power operation and maintenance personnel.

[0057] The task entity is used to represent the type of work activity currently being performed by the operation and maintenance personnel, and is associated with a task attribute set, which includes the task duration and the task risk level.

[0058] The water-fall risk status entity is used to characterize the risk evolution status formed by the combined effects of personnel status, environmental factors and task status under specific time and space conditions, including normal status, risk initiation status, risk accumulation status, high-risk status and near-water-fall status.

[0059] This application also provides a device for early warning of the risk of offshore wind power maintenance personnel falling overboard, including:

[0060] The fusion module is used to generate multi-source fused water-fall risk data based on the status information of maintenance personnel, spatial status information, and environmental and task status information.

[0061] The acquisition module is used to acquire a pre-constructed initial knowledge representation of the risk of falling into the water. The initial knowledge representation of the risk of falling into the water is pre-constructed based on historical risk event data and domain prior knowledge, and includes multiple key entities that affect the risk of falling into the water and the static correlation between each key entity.

[0062] The update module is used to update the attributes of the key entity based on the multi-source fused water risk data. When the updated attributes meet the preset risk state triggering conditions, the corresponding attributes are converted into real-time water risk facts with time stamps.

[0063] A module is established to establish a temporal correlation between the real-time water risk facts based on their occurrence time and spatial location.

[0064] The construction module is used to integrate the key entities, the static relationships, the real-time fall-in-water risk facts and their time information to construct a time-series fall-in-water risk knowledge graph;

[0065] The early warning module is used to reason about the evolution trend of falling water risk under the combined effect of multiple factors based on the historical evolution pattern in the time-series falling water risk knowledge graph and the real-time falling water risk facts, generate and output interpretable falling water risk early warning information, and the interpretable falling water risk early warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status association path.

[0066] The beneficial effects of this invention are as follows:

[0067] This method integrates multi-source fall-in-water risk data to construct a knowledge graph containing key entities such as personnel, spatial location, environmental factors, and operational tasks, as well as their temporal relationships. Using this knowledge graph, multi-factor temporal correlation reasoning is performed to achieve accurate prediction of the evolution trend of fall-in-water risk and the generation of interpretable early warnings. This solves the black-box problem of existing pure data-driven methods and the false alarm and missed alarm problems caused by insufficient characterization of multi-factor temporal correlation. Attached Figure Description

[0068] Figure 1 This is a flowchart illustrating the method of an embodiment of this application;

[0069] Figure 2 This is a structural block diagram of the device in the embodiments of this application. Detailed Implementation

[0070] Reference Figure 1 This application provides a method for early warning of the risk of offshore wind power maintenance personnel falling into the water, including:

[0071] S101, based on the status information of maintenance personnel, spatial status information, and environmental and task status information, forms multi-source fusion data on the risk of falling into water;

[0072] S102, Obtain the pre-constructed initial knowledge representation of the risk of falling into the water. The initial knowledge representation of the risk of falling into the water is pre-constructed based on historical risk event data and domain prior knowledge, and includes multiple key entities that affect the risk of falling into the water and the static relationship between each key entity.

[0073] S103, based on multi-source fusion of water risk data, update the attributes of key entities. When the updated attributes meet the preset risk state triggering conditions, convert the corresponding attributes into real-time water risk facts with time stamps.

[0074] S104, Based on the occurrence time and spatial location of each real-time fall-in-water risk fact, establish the temporal correlation between each real-time fall-in-water risk fact;

[0075] S105 integrates key entities, static relationships, real-time facts about the risk of falling into the water and their time information to construct a time-series knowledge graph of the risk of falling into the water.

[0076] S106. Based on the historical evolution patterns and real-time facts of falling into water risk in the time-series knowledge graph, reason about the evolution trend of falling into water risk under the combined effect of multiple factors, generate and output interpretable falling into water risk warning information. The interpretable falling into water risk warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status association paths.

[0077] In step S101, the status information of the maintenance personnel is obtained through a wearable safety monitoring device for offshore wind power maintenance personnel. The status information of the maintenance personnel includes their vital signs and behavioral status information. Among them, the vital signs information includes heart rate, body temperature, and blood pressure, which are used to reflect the physiological state of the maintenance personnel. Abnormal fluctuations in heart rate or body temperature may indicate that the personnel are unwell or under stress. The behavioral status information includes triaxial acceleration, motion velocity, and attitude angle, which are used to reflect the body posture and motion state of the maintenance personnel. Abnormal changes in attitude angle may indicate that the personnel are experiencing posture instability.

[0078] The spatial status information of maintenance personnel is also obtained through wearable safety monitoring devices for offshore wind power maintenance personnel. These devices integrate GPS or BeiDou positioning modules. The spatial status information includes the maintenance personnel's current location coordinates, movement trajectory, and spatial movement status information. The current location coordinates are used to determine the precise spatial location of the maintenance personnel in areas such as the wind turbine tower, nacelle, ship deck, or platform edge; the movement trajectory records the spatial movement path of the maintenance personnel within a certain time window, reflecting whether they are gradually approaching a danger zone; and the spatial movement status information includes movement speed and direction.

[0079] Environmental and mission status information is obtained through the offshore wind power operation and maintenance management system, which includes a SCADA system, a meteorological monitoring station, and an operation and maintenance work order system.

[0080] Environmental status information includes wind speed, wind direction, temperature, and visibility, used to characterize the safety conditions of the external environment in which maintenance personnel are located. Among these, wind speed is a key factor affecting the stability of personnel operations. Task status information includes task type, task area, task stage, and task duration, used to characterize the maintenance activities currently being performed by maintenance personnel. The task stage (such as tower climbing stage, cabin operation stage, deck operation stage) reflects the specific work phase that personnel are currently in, and the task duration reflects the length of time that personnel have been working continuously.

[0081] After obtaining the above status information, spatial status information, and environmental and task status information, missing value imputation, outlier detection, and time synchronization are performed on these three types of data. The three types of data are then correlated and integrated according to a unified timestamp to form multi-source fused water-fall risk data X(t), represented as:

[0082] In the formula, t represents time, H(t) represents the status data of operation and maintenance personnel, L(t) represents the spatial status data of operation and maintenance personnel, and E(t) represents the environmental and task status data.

[0083] In step S102, based on historical risk event data and prior domain knowledge, the key entity types affecting the risk of offshore wind power operation and maintenance personnel falling into the water are identified, including personnel entity types, spatial entity types, environmental entity types, task entity types, and the risk of falling into the water status entity type.

[0084] The "Personnel" entity type represents the category of personnel involved in offshore wind power operation and maintenance activities. In the initial knowledge representation, this entity type predefines personnel status attributes, including vital signs and behavioral attributes. The vital signs attributes correspond to heart rate, body temperature, and blood pressure, while the behavioral attributes correspond to triaxial acceleration, velocity, and attitude angle. In the subsequent real-time operation phase, this entity type will be instantiated into specific individual personnel, and their attributes will be filled with real-time collected values.

[0085] Spatial entity types are used to characterize the physical space categories involved in offshore wind power operation and maintenance. These entity types predefine spatial risk attributes, including hazard levels. Specific spaces corresponding to a spatial entity type include the ship deck area, tower interior area, engine room area, platform edge area, and pre-defined hazardous operation areas. During subsequent real-time operation, this entity type will be instantiated into specific spatial area instances.

[0086] The environmental entity type is used to characterize categories of external environmental factors that affect the safety of offshore wind power operation and maintenance personnel. This entity type predefines environmental state attributes, corresponding to wind speed, wind direction, temperature, and visibility. In the subsequent real-time operation phase, each environmental factor will be instantiated into specific environmental parameter instances.

[0087] The task entity type represents the category of operational activities currently being performed by maintenance personnel. This entity type predefines task attributes, including task duration and task risk level. Specific tasks corresponding to this task entity type include blade maintenance tasks, tower maintenance tasks, and offshore lifting tasks. During subsequent real-time operation, this entity type will be instantiated into specific task instances.

[0088] The "fall into water risk state" entity type is used to characterize the risk evolution state category formed by the combined effects of personnel status, environmental factors, and task status under specific time and space conditions. This entity type predefines risk state attributes, and the corresponding risk states include normal state, risk initiation state, risk accumulation state, high-risk state, and near-fall into water state. In the subsequent real-time operation phase, this entity type will be instantiated into specific risk state instances.

[0089] Based on prior domain knowledge, static association types are defined for each entity type. Among them, the personnel-space relationship is used to characterize the spatial association between personnel entities and spatial entities, including two relationship types: located and close. Located indicates the space where the personnel are currently located; close indicates that the distance between the personnel and a specific space is less than a preset safety threshold.

[0090] The space-fall-into-water risk state relationship is used to characterize the risk association between a spatial entity and a risk state entity, including two relationship types: easy to generate and associated with. Easy to generate indicates that the space is likely to induce a corresponding risk state under preset conditions; associated with indicates that there is an inherent correlation between the space and a specific risk state.

[0091] The environment-drowning risk state relationship is used to characterize the risk-driven relationship between environmental entities and risk state entities. This includes adding relationship types that indicate that the environmental factor increases the probability of a specific risk state occurring.

[0092] The task-risk status relationship is used to characterize the risk association between a task entity and a risk status entity, including the association type, which indicates that there is a correlation between the task type and a specific risk status.

[0093] Based on historical risk event data, expert experience, and offshore wind power operation and maintenance safety standards, the impact of each correlation on the formation process of the risk of falling overboard is quantitatively analyzed, and an initial risk correlation weight ω is assigned to each static correlation. ij .

[0094] Finally, based on the aforementioned definitions and the generated entity types, static relationship types, and initial risk association weights, an initial knowledge representation of the risk of falling into the water is formed. This knowledge representation provides a static foundational framework for constructing a temporal knowledge graph of the risk of falling into the water during the subsequent real-time operation phase.

[0095] In the subsequent real-time operation phase, the entity types in this initial knowledge representation will be instantiated into concrete entity instances. For example, the personnel entity type is instantiated into the specific personnel object "Operation and Maintenance Personnel A," and its predefined attribute items are filled with real-time collected values ​​(such as heart rate 72 bpm, normal posture angle, etc.). Real-time water fall risk facts are generated based on these instantiated entities and their attribute values ​​and dynamically added to the knowledge graph, forming a temporal knowledge graph that evolves over time.

[0096] In this embodiment of the application, step S103 includes:

[0097] S1031, based on the status information of maintenance personnel in the multi-source fusion water risk data, update the numerical values ​​of the personnel status attribute set associated with the personnel entity; the personnel status attribute set includes vital sign status attributes and behavioral status attributes, and record the time stamp of this update.

[0098] For example: Assume the current time is t1. A wearable device collects the heart rate of maintenance personnel A, which is 118 bpm, within the normal range of 60 to 100 bpm. Simultaneously, the posture angle is collected as 35 degrees, within the normal tilt threshold of 15 degrees. Based on the maintenance personnel status information from the multi-source fusion water-fall risk data, the personnel status data for this current moment is extracted. The personnel status attribute set associated with the personnel entity is then updated according to the corresponding attribute items, i.e., the heart rate attribute is updated to 118 bpm and the posture angle attribute is updated to 35 degrees. The time stamp of this update is recorded as t1.

[0099] S1032, based on the spatial state information in the multi-source fusion water risk data, update the spatial risk attribute set associated with the spatial entity numerically, and record the time stamp of this update.

[0100] Assuming the current time is t2, the distance between the current location coordinates of maintenance personnel A and the edge area of ​​the platform, collected by the wearable device positioning module, is 0.5 meters, with a preset safety threshold of 1 meter. Based on the spatial state information in the multi-source fusion water-fall risk data, the spatial state data for the current time is extracted, and the spatial risk attribute set associated with the spatial entity is updated according to the corresponding attribute items. That is, the distance attribute from the platform edge is updated to 0.5 meters, and the time identifier of this update is recorded as t2.

[0101] S1033, based on the environmental and task status information in the multi-source fusion water risk data, update the values ​​of the environmental status attribute set associated with the environmental entity and the task attribute set associated with the task entity, and record the time stamp of this update.

[0102] Assuming the current time is t3, the wind speed collected by the operation and maintenance management system is 15 meters per second, and the preset safety threshold is 10 meters per second. Based on the environmental state information in the multi-source fusion water-fall risk data, the environmental state data for the current time is extracted, and the environmental state attribute set associated with the environmental entity is updated according to the corresponding attribute items. That is, the wind speed attribute is updated to 15 meters per second, and the time identifier of this update is recorded as t3.

[0103] Assuming the current time is t4, the operation and maintenance management system has collected data showing that maintenance personnel A is currently performing an offshore lifting operation, which has been ongoing for 3 hours, with a preset safety time of 2 hours. Based on the task status information from the multi-source fusion data on the risk of falling overboard, the task status data for the current time is extracted. The task attribute set associated with the task entity is then updated according to the corresponding attribute items, specifically updating the task type to offshore lifting and the task duration to 3 hours. The time stamp of this update is recorded as t4.

[0104] S1034, Based on the preset risk status determination rules, analyze the updated attributes of each key entity. When the attribute of any key entity or the relationship between entities meets the preset risk status triggering condition, convert the corresponding attribute or relationship into a real-time drowning risk fact with a time stamp, represented as: ;

[0105] In the formula, Indicates a risk entity. Indicates the relationship between entities. Indicates related entities or attribute values. Indicates the time when the risk occurred.

[0106] The real-time risk of falling into water includes changes in personnel status, spatial status, environmental status, and task status. For example, when a personnel entity's heart rate attribute is updated and exceeds a preset threshold, a personnel status change fact is generated, with the risk entity being the personnel entity, the association being "abnormal status," the associated attribute being the heart rate value, and the occurrence time being the current moment. Similarly, when a spatial entity's distance from the platform edge is updated and falls below a preset safety threshold, a spatial status change fact is generated, with the risk entity being the personnel entity, the association being "approaching," the associated entity being the platform edge area, and the occurrence time being the current moment. Furthermore, when an environmental entity's wind speed attribute is updated and exceeds a preset threshold, an environmental status change fact is generated, with the risk entity being the environmental entity, the association being "wind speed exceeding the limit," the associated attribute being the wind speed value, and the occurrence time being the current moment.

[0107] For example, according to the preset risk status judgment rules, when the heart rate exceeds 100 bpm and the posture angle exceeds 15 degrees, a personnel abnormal status risk status is triggered. The current heart rate of 118 bpm exceeds 100 bpm, and the posture angle of 35 degrees exceeds 15 degrees; both conditions are met. Therefore, this personnel abnormal status is converted into a real-time water-fall risk fact with a time stamp. That is, the risk entity is personnel entity A, the association relationship is abnormal status, the associated attribute values ​​are heart rate 118 bpm and posture angle 35 degrees, and the occurrence time is t1.

[0108] According to the preset risk status determination rules, when the distance between a person and the edge area of ​​the platform is less than 1 meter, the risk status of approaching a dangerous area is triggered. The current distance is 0.5 meters, which is less than 1 meter, thus meeting the condition. Therefore, the approaching dangerous area is converted into a real-time fall into water risk fact with a time stamp, that is, the risk entity is the person entity A, the relationship is approach, the associated entity is the platform edge area, and the occurrence time is t2.

[0109] According to the preset risk status determination rules, a severe environmental condition risk status is triggered when the wind speed exceeds 10 meters per second. The current wind speed of 15 meters per second exceeds 10 meters per second, thus meeting the condition. Therefore, this severe environmental condition is converted into a real-time water-falling risk fact with a time stamp. That is, the risk entity is an environmental entity, the correlation is wind speed exceeding the standard, the correlation attribute value is 15 meters per second, and the occurrence time is t3.

[0110] According to the preset risk status determination rules, when the duration of offshore hoisting operations exceeds 2 hours, a task fatigue risk status is triggered. The current duration of 3 hours has exceeded 2 hours, thus meeting the condition. Therefore, the task fatigue risk is converted into a real-time risk of falling into the water with a time stamp. That is, the risk entity is the task entity, the associated relationship is the offshore hoisting task, the associated attribute value is the duration of 3 hours, and the occurrence time is t4.

[0111] Through the above process, all key entities that meet the preset risk state triggering conditions during the attribute update process are converted into corresponding real-time drowning risk facts, providing a data foundation for establishing the temporal correlation between the facts.

[0112] In this embodiment of the application, step S104 includes:

[0113] S1041, obtain the entity type, relationship type, occurrence time and spatial location corresponding to each real-time fall-in-water risk fact.

[0114] Suppose that the following three real-time facts regarding the risk of falling into the water have been generated at the current moment:

[0115] Fact A: The fact of change in personnel status, occurring at time t. A =09:00:05, spatial location is p A =(10,20).

[0116] Fact B: The environmental state changed at time t. B =09:00:10, spatial location is p B =(30,25).

[0117] Fact C: A change in spatial state, occurring at time t. C =09:00:15, spatial location is p C =(35,28).

[0118] S1042, based on the occurrence time t of any two real-time fall-in-water risk facts. i and t j The time correlation coefficient is calculated using the time decay function. :

[0119]

[0120] Where k is the preset time decay coefficient and the time correlation coefficient. Used to characterize the time dependency in the continuous evolution of risk states;

[0121] For facts A and B, the time interval is:

[0122] With a preset time decay coefficient k=0.1, the time correlation coefficient is:

[0123]

[0124] For facts B and C, the time interval is:

[0125] The time correlation coefficient is:

[0126]

[0127] For facts A and C, the time interval is:

[0128] The time correlation coefficient is:

[0129]

[0130] S1043, based on the spatial location p corresponding to any two real-time fall-in-water risk facts i and p j Calculate the spatial correlation coefficient :

[0131]

[0132] Spatial correlation coefficient It is used to characterize the spatial constraint relationship between changes in the area of ​​human activity and the danger zone.

[0133] The spatial locations of facts A and B are pA=(10,20) and pB=(30,25) respectively, and the spatial distance is:

[0134]

[0135] The spatial correlation coefficient between fact A and fact B is:

[0136] Similarly, the spatial correlation coefficient between fact B and fact C is: The spatial correlation coefficient between fact A and fact C is:

[0137] S1044, based on the time correlation coefficient Spatial correlation coefficient Calculate the overall correlation strength between any two real-time facts of the risk of falling into the water. : ;

[0138] in, The first preset weighting coefficient, The second preset weighting coefficient, This is the offset of the overall association strength for the preset temporal association relationships.

[0139] Let the first preset weighting coefficient be... =0.4, second preset weighting coefficient =0.6, offset =0. Therefore, we can obtain:

[0140] The spatial correlation coefficient between fact A and fact B is:

[0141] The spatial correlation coefficient between fact B and fact C is:

[0142] The spatial correlation coefficient between fact A and fact C is:

[0143] S1045, when the overall correlation strength When the risk exceeds a preset threshold, a temporal correlation is established between two real-time fall-in-water risk facts, and the overall correlation strength is considered. As the association weight of the corresponding temporal relationship.

[0144] Let the preset correlation threshold be θ=0.3.

[0145] because =0.2704<0.3, therefore, no temporal relationship is established between fact A and fact B.

[0146] because =0.3304>0.3, establish a temporal association between fact B and fact C, with an association weight of 0.3304.

[0147] because =0.1692<0.3, therefore, no temporal relationship is established between fact A and fact C.

[0148] In the example above, facts B and C are associated in a temporal sequence. Fact B is an environmental state change, occurring at 09:00:10, and located in the external area of ​​the wind turbine tower; fact C is a spatial state change, occurring at 09:00:15, and located in the edge area of ​​the platform. The time interval between the two is 5 seconds, and their spatial distance is relatively short. The overall association strength is 0.3304, exceeding the preset threshold of 0.3.

[0149] The establishment of this temporal correlation indicates that the wind speed exceeding the limit event occurred before the event of approaching the platform edge. The two events are 5 seconds apart in time and are spatially close. The wind speed exceeding the limit may have prompted maintenance personnel to move towards the platform edge. Together, they constitute a continuous risk evolution path. In the temporal knowledge graph, this correlation is stored in the form of directed edges, with edge attributes including a time interval of 5 seconds and a correlation weight of 0.3304.

[0150] Although the time interval between facts A and B is also 5 seconds, the spatial distance is relatively large, and the overall correlation strength is 0.2704, which is lower than the preset threshold. Therefore, no correlation is established, indicating that spatial distance is an important constraint affecting the establishment of correlation between risk factors.

[0151] In step S105, key entities, static relationships, real-time facts about the risk of falling into the water and their time information are integrated to construct a time-series knowledge graph of the risk of falling into the water.

[0152] The time-series knowledge graph of drowning risk uses key entities as nodes, static and time-series relationships as edges, and the occurrence time of each risk fact and the corresponding relationship weight as edge attributes for storage.

[0153] Key entities serve as nodes in the knowledge graph, with each node associated with a corresponding entity type and its current attribute value. Entity types include personnel, spatial, environmental, task, and drowning risk status entities. Node attribute values ​​include the real-time status data corresponding to each entity.

[0154] Static relationships, as static edges in a knowledge graph, are used to represent the inherent relationship types between key entities. Each static edge has an initial risk association weight as an edge attribute. Static relationships include personnel-space relationships, space-fall-over-risk-state relationships, environment-fall-over-risk-state relationships, and task-fall-over-risk-state relationships.

[0155] Temporal relationships, as temporal edges in the knowledge graph, are used to represent the evolutionary relationships between real-time water-fall risk facts in the temporal and spatial dimensions. Each temporal edge has a comprehensive association strength as an edge attribute and a corresponding risk fact occurrence time as a timestamp attribute.

[0156] The knowledge graph representation of time-series water-fall risk is as follows:

[0157] in, This indicates the head node corresponding to the risky entity. This indicates the type of association between the head node and the tail node. Indicates the tail node or attribute value. This indicates the time when the risk event corresponding to this relationship occurred. This indicates the association weight corresponding to the association relationship; when When it is a static association, The value is taken as the initial risk association weight; when When it is a temporal relationship, The value is determined by the overall correlation strength.

[0158] The time-series knowledge graph of water-related risks provides a structured data foundation for risk evolution pattern matching and multi-factor time-series correlation reasoning in subsequent steps.

[0159] In this embodiment of the application, step S106 includes:

[0160] S1061, based on the real-time risk of falling into water at the current moment, match the corresponding risk evolution pattern in the time-series risk knowledge graph to determine the evolution trend of the current risk status of maintenance personnel.

[0161] Based on the real-time risk of falling into the water at the current moment, obtain the corresponding risk entity type, entity relationship type, event occurrence time, and risk relationship weight. Among them, the risk entity type includes personnel entity, spatial entity, environmental entity, mission entity, and falling into the water risk status entity; the entity relationship type includes being close enough to generate a relationship and increasing the relationship.

[0162] Using risk entity type and entity relationship type as search criteria, historical risk facts with the same risk entity type and entity relationship type as the current risk state are extracted from the time-series risk knowledge graph and arranged chronologically to form a historical risk fact sequence. The retrieved historical risk facts are then arranged in chronological order of occurrence, forming several historical risk fact sequences. Each historical risk fact sequence corresponds to a candidate risk evolution pattern. The candidate risk evolution pattern is represented as follows: .

[0163] in Indicates the entity type corresponding to the i-th historical risk fact. This indicates the type of entity association corresponding to the i-th historical risk fact. represents the time when the i-th historical risk fact occurred, m represents the number of facts in this evolutionary pattern, and n represents the sequence number of the candidate evolutionary pattern.

[0164] Calculate the similarity between the temporal features corresponding to the current risk state and the temporal features corresponding to each candidate risk evolution pattern in the historical risk fact sequence.

[0165] The time-series features corresponding to the current risk state are represented in vector form. Where d represents the feature dimension, The dimensions include entity type encoding, relationship type encoding, risk association weight, and time interval features.

[0166] The temporal features corresponding to each candidate risk evolution pattern in the historical risk fact sequence are extracted into vector form. , where n represents the sequence number of the candidate evolutionary mode.

[0167] The extraction methods for time-series features include: mapping entity types to corresponding type encoding values, mapping relationship types to corresponding relationship encoding values, using risk association weights as numerical features, and using time intervals between facts as time-series features.

[0168] Calculate the time series feature X corresponding to the current risk state. t Temporal features X corresponding to each candidate risk evolution mode n The similarity between them is calculated using cosine similarity:

[0169]

[0170] Where d represents the feature dimension. This represents the value of the current time-series feature vector in the q-th dimension. This represents the value of the temporal feature vector of the nth candidate evolutionary pattern in dimension q. The similarity value ranges from -1 to 1, with the value closer to 1 indicating a greater similarity in temporal features between the current risk state and the candidate evolutionary pattern.

[0171] The candidate risk evolution pattern with the highest similarity is identified as the matching risk evolution pattern. ,Right now: Based on the matched risk evolution pattern, the evolution trend corresponding to the current risk status of the operations and maintenance personnel is determined. Specifically, the sequence of subsequent facts following the events that have already occurred in the matched risk evolution pattern is obtained, and the direction of state change corresponding to the subsequent fact sequence is taken as the evolution trend of the current risk status of the operations and maintenance personnel. For example, if the matched risk evolution pattern is excessive wind speed → approaching the platform edge → attitude imbalance → high-risk state, and the current fact is approaching the platform edge, then the evolution trend is to continue developing towards attitude imbalance → high-risk state.

[0172] Assuming the current time is 9:00 AM, the generated real-time risk of falling into the water is fact C, which is the spatial state change fact. The risk entity is maintenance personnel A, the relationship is proximity, the associated entity is the platform edge area, the time of occurrence is 9:00 AM, and the spatial location is the platform edge.

[0173] Based on the current real-time risk of falling into the water, and using personnel entities and proximity relationships as search criteria, historical risk fact sequences are retrieved from the temporal knowledge graph of falling into the water risk. Assume the graph stores the following three historical evolution patterns:

[0174] Historical pattern 1 involves wind speeds exceeding the standard, approaching the platform edge, causing personnel to lose their balance, and ultimately entering a high-risk state. The evolution path length is 4.

[0175] Historical pattern two involves the wind speed exceeding the limit and approaching the edge of the platform, then returning to the inside of the tower, eventually eliminating the risk. The evolution path length is 4.

[0176] Historical pattern three involves an abnormal heart rate followed by approaching the edge of the platform, then loss of posture, and finally entering a high-risk state, with an evolutionary path length of 4.

[0177] The similarity between the current risk state's temporal characteristics and each candidate historical pattern is calculated. Matching is performed based on entity type, relationship type, and occurrence time sequence. The current state sequence, approaching the platform edge, is identical to the starting sequences of historical patterns one and two, but differs from historical pattern three. Further considering other real-time risk facts present at the current moment, namely fact B (exceeding wind speed limits, occurring at 8:55 AM), the current state is determined to have the highest match with historical pattern one. Historical pattern one is identified as the matching risk evolution pattern, thus determining the evolution trend of the current risk state of maintenance personnel A as follows: after exceeding wind speed limits, approaching the platform edge, then experiencing imbalance, and finally continuing to evolve towards higher risk.

[0178] S1062, based on the evolution trend, obtain the temporal correlation between key entities in the temporal knowledge graph of drowning risk, disseminate information on different risk factors along the temporal correlation path, obtain comprehensive risk correlation characteristics, and obtain the correlation strength of the evolution from the current risk state to the target risk state based on the comprehensive risk correlation characteristics.

[0179] Specifically, step S1062 includes:

[0180] Based on the evolution trend, the key entities involved in risk evolution are identified, including personnel entities, space entities, environmental entities, mission entities, and risk state entities. Specifically, all entity types involved in the matched risk evolution pattern are obtained, including personnel entities, space entities, environmental entities, mission entities, and fall-over risk state entities, and the instantiated objects of these entities at the current moment and their current attribute values ​​are obtained from the temporal fall-over risk knowledge graph. For example, when the matched evolution pattern is excessive wind speed → approaching the platform edge → attitude imbalance → high-risk state, the key entities involved in risk evolution include the wind speed entity in the environmental entities, the platform edge area entity in the space entities, maintenance personnel A in the personnel entities, and the high-risk state entity in the risk state entities.

[0181] Obtain the temporal relationships and corresponding temporal association paths among the key entities involved in the risk evolution in the temporal knowledge graph of drowning risk. The temporal association path is a directed path that connects the risk facts in chronological order of occurrence. Each edge in the path represents a temporal association, and each edge has an association weight and time interval attribute.

[0182] Taking the evolutionary path of excessive wind speed → approaching the platform edge → attitude imbalance → high-risk state as an example, it includes the following temporal correlations:

[0183] The wind speed exceeded the standard at 8:55 AM, and the event near the platform edge occurred at 9:00 AM. The correlation weight was 0.3304, and the time interval was 5 minutes.

[0184] The event of approaching the platform edge occurred at 9:00 AM, and the event of posture imbalance occurred at 9:05 AM, with a correlation weight of 0.2758 and a time interval of 5 minutes.

[0185] The postural imbalance occurred at 9:05 AM, which is related to the high-risk management event that occurred at 9:10 AM, with a correlation weight of 0.4102 and a time interval of 5 minutes.

[0186] Along the temporal correlation path, the risk factor characteristics of each key entity are sequentially passed to adjacent entities. Specifically, the propagation method of risk factor characteristics is as follows: the starting entity multiplies its risk factor characteristics by the correlation weight of its temporal correlation relationship with the next entity, and passes the product to the next entity as the risk propagation value received by that entity. When any entity receives risk factor characteristics transmitted from multiple adjacent entities, they are weighted and summed according to the risk correlation weights corresponding to each correlation relationship to form the comprehensive risk correlation characteristics of that entity.

[0187] The above evolutionary path will be used as an example for illustration:

[0188] The starting entity is an environmental entity, whose risk factor characteristic is an initial risk contribution value of 0.6 corresponding to a wind speed of 12 m / s. The environmental entity multiplies its risk factor characteristic of 0.6 by the association weight of 0.3304 with the spatial entity, resulting in 0.1982, and transmits it to the spatial entity. The spatial entity itself has an inherent risk attribute of being located at the edge of the platform, with a high hazard level and an inherent risk contribution value of 0.5. The spatial entity receives only the risk factor characteristic transmitted from one neighboring entity. This received value of 0.1982 is aggregated with its own inherent risk contribution value of 0.5 to obtain the spatial entity's comprehensive risk association characteristic of 0.6982.

[0189] The spatial entity multiplies its comprehensive risk correlation feature (0.6982) by the correlation weight (0.2758) with the personnel entity, resulting in 0.1925, and transmits this value to the personnel entity. The personnel entity's current state feature is a heart rate increased to 118 beats per minute, with a risk contribution value of 0.4. The personnel entity receives risk factor features transmitted from only one neighboring entity, and aggregates this received value (0.1925) with its own state risk contribution value (0.4) to obtain the personnel entity's comprehensive risk correlation feature of 0.5925.

[0190] If an entity simultaneously receives risk factor characteristics transmitted from multiple adjacent entities, a weighted sum is calculated according to the association weights corresponding to each relationship. For example, if a risk state entity simultaneously receives a risk propagation value of 0.1982 from an environmental entity, 0.1925 from a spatial entity, and 0.1800 from a personnel entity, the weighted sum would be 0.1982 plus 0.1925 plus 0.1800, equaling 0.5707. This sum is then aggregated with the entity's inherent risk characteristics to form the entity's comprehensive risk association characteristics.

[0191] The mathematical expression for the propagation and aggregation process is: for entity ej, the propagation and aggregation process it receives from neighboring entity ej is as follows: i The risk propagation value is V i ⋅ω ij V i For entity e i Risk characteristic value, ω ij For entity e i With e j The weight of the temporal relationship between entities. Entity e j The comprehensive risk association characteristics are as follows:

[0192]

[0193] in, Representing entities Its inherent risk characteristics Indicates to entity The set of adjacent entities that transmit risk factors, where f(⋅) represents the aggregation function.

[0194] Based on the comprehensive risk correlation characteristics, we analyze the trend of risk status change under the combined effect of different risk factors, and obtain the correlation strength of the evolution of the current risk status to the target risk status.

[0195] Specifically, the comprehensive risk association features corresponding to the target risk state entity in the time-series knowledge graph of drowning risk are obtained, and these comprehensive risk association features are used as the association strength for the evolution of the current risk state to the target risk state. The target risk state is the next risk state after the current moment in the matched risk evolution pattern, and the association strength is used to characterize the probability that the risk state will develop into the target risk state along the current evolution trend under the combined effect of multiple factors.

[0196] Taking the above example, the target risk state is a high-risk state, and its corresponding comprehensive risk correlation feature is 0.5925, that is, the correlation strength of the current risk state evolving into a high-risk state is 0.5925.

[0197] The above process propagates risk factor characteristics step by step along the temporal correlation path and performs weighted aggregation when multiple risk factors converge, realizing the trend analysis of risk status changes under the combined effect of multiple factors. The obtained correlation strength provides a quantitative basis for subsequent calculation of the probability of falling into the water.

[0198] S1063, based on the correlation strength, integrates the current risk status characteristics with the comprehensive risk correlation characteristics, inputs them into the fall-in-water risk prediction model, and calculates the probability that maintenance personnel will enter the target fall-in-water risk state within a preset future time range.

[0199] In step S1063, the current risk state feature is the encoded feature vector corresponding to the current risk state extracted from the time-series drowning risk knowledge graph, represented as follows: , where m represents the dimension of the current risk status characteristics. The current risk status characteristics include the entity type code, entity attribute value code, and risk association weight of each key entity at the current moment.

[0200] The comprehensive risk association feature is the comprehensive risk association feature value corresponding to the target risk state entity obtained in step S1062. .

[0201] Characteristics of the current risk status Characteristic values ​​associated with comprehensive risk The features are concatenated and merged to form a fused feature vector: .

[0202] For example, the current risk status features are [0.8, 0.7, 0.2], the comprehensive risk association feature values ​​are 0.5925 and 0.5925, and the fused feature vector obtained after concatenation is [0.8, 0.7, 0.2, 0.5925].

[0203] fuse feature vectors The data is input into a pre-built water risk prediction model to calculate the probability that maintenance personnel will enter a target water risk state within a preset future time range.

[0204] The drowning risk prediction model is pre-built based on the embedded features of the time-series knowledge graph, specifically a logistic regression model or a neural network model. It is trained by supervised learning, using historical fusion feature vectors as training samples and historical actual risk state changes as training labels.

[0205] The drowning risk prediction model is expressed as follows:

[0206]

[0207] Where W represents the weight vector of the drowning risk prediction model, b represents the bias term of the drowning risk prediction model, and σ(⋅) represents the Sigmoid activation function. This indicates the probability that maintenance personnel will enter a target water-fall risk state within a preset future time frame.

[0208] Taking the example in step S1062, the fused feature vector is [0.8, 0.7, 0.2, 0.5925], the model weight vector is [0.5, 0.3, 0.2, 0.4], and the bias term is −0.3, then: +b=0.587; That is, the probability that maintenance personnel A will enter a high-risk state within the next 5 minutes is approximately 64.27%.

[0209] S1064. Based on the contribution of each key entity to the probability, identify the key entities that affect the change of the current risk state, extract the corresponding risk state association paths, generate interpretable drowning risk warning information and output it.

[0210] In this embodiment of the application, step S1064 includes:

[0211] In this embodiment, a contribution attribution method is used to calculate the contribution of each key entity in the time-series knowledge graph of drowning risk to the probability. The contribution attribution method is a model weight-based attribution method. For each feature dimension in the input features of the drowning risk prediction model, the corresponding model weight coefficient is multiplied by the value of that feature dimension, and the resulting product is taken as the contribution value of that feature dimension to the prediction probability.

[0212] For key entities Its contribution This is the sum of the contribution values ​​of all feature dimensions corresponding to this entity: .

[0213] in, Represents key entity e k The set of feature dimension indices corresponding to the fused feature vectors This represents the model weight coefficient corresponding to the q-th feature dimension. This represents the value of the q-th feature dimension.

[0214] For example, for a wind speed entity, its corresponding feature dimensions include a wind speed type encoding dimension and a wind speed attribute value dimension. If the wind speed type encoding feature value is 0.8, the corresponding weight is 0.5, and the contribution value is 0.4; if the wind speed attribute value feature value is 0.6, the corresponding weight is 0.3, and the contribution value is 0.18. Then the total contribution of the wind speed entity is 0.58.

[0215] The key entities are ranked according to their contribution level, and those whose contribution level exceeds a preset threshold are identified as key entities that affect the current risk status.

[0216] Specifically, the contributions of each key entity are sorted in descending order to obtain a contribution ranking list. Key entities with a contribution greater than or equal to a preset contribution threshold are marked as key causative entities.

[0217] For example, the calculated contributions of each entity are as follows: wind speed entity contribution is 0.40, spatial location entity contribution is 0.21, personnel status entity contribution is 0.04, and comprehensive risk association feature contribution is 0.237. With a preset threshold of 0.1, the key entities whose contributions exceed the threshold are, in order, wind speed entity, entities corresponding to comprehensive risk association features, and spatial location entity. The contribution of personnel status entity is below the threshold and is not considered a key causative factor.

[0218] Based on key entities, corresponding risk facts, and relationships between entities, the risk state association paths experienced from the initial risk state to the current risk state are extracted from the time-series water-fall risk knowledge graph.

[0219] The initial risk state is the risk state at the start time in the matched risk evolution pattern, and the current risk state is the risk state at the current time. The risk state association path starts from the initial risk state and ends at the current risk state, including all key entities on the path, their temporal relationships, and corresponding risk facts.

[0220] Generate and output interpretable fall-in-water risk warning information. The interpretable fall-in-water risk warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status correlation path.

[0221] The risk level is determined based on the probability range to which the probability value calculated in step S1053 belongs. A probability greater than 60% is considered high risk, a probability between 30% and 60% is considered medium risk, and a probability less than 30% is considered low risk.

[0222] The risk factors and their contribution ranking include all key entities whose contribution exceeds a preset threshold and their corresponding contribution, arranged from highest to lowest contribution.

[0223] Risk facts include all key risk facts and the timing of their occurrence during the evolution from the initial risk state to the current risk state.

[0224] The risk state association path includes all intermediate risk states experienced from the initial risk state to the current risk state and their corresponding temporal evolution directions. For example, the generated interpretable fall-in-water risk warning information is as follows: the risk level is high risk, with a risk probability of 64.27%; the risk factors and their contribution ranking are: wind speed 45.14%, spatial location platform edge area 23.70%, and personnel heart rate status 4.51%; the risk facts are: heart rate increased to 118 beats per minute at 08:50, wind speed reached 12 meters per second at 08:55, and approaching the platform edge area at 09:00; the risk state association path is: after the wind speed exceeds the standard, approaching the platform edge area, and then entering a high-risk state.

[0225] In this embodiment of the application, the method further includes:

[0226] S107, Based on the deviation between the output interpretable water risk warning information and the actual risk status change results, update the association weights between key entities and water risk status entities in the time-series water risk knowledge graph.

[0227] S108. Based on the updated relationships and newly added real-time water-falling risk facts, the key entities, relationships, and risk evolution patterns in the time-series water-falling risk knowledge graph are updated to optimize the time-series water-falling risk knowledge graph.

[0228] After outputting interpretable fall-in-water risk warning information, the system continuously collects the actual risk status changes at subsequent times. The actual risk status changes refer to the actual changes in risk status that occur within a preset time window after the warning information is output, including whether a fall-in-water risk event actually occurs, the trajectory of the actual risk level change, and the actual change data of the attributes of each key entity.

[0229] The actual changes in risk status are compared with the predicted results in the output early warning information to calculate the prediction deviation. The prediction deviation includes the prediction deviation of risk level, the prediction deviation of risk occurrence time, and the identification deviation of key causal factors.

[0230] When the prediction deviation exceeds a preset deviation threshold, the association weight update mechanism is triggered. Specifically, for each static association between personnel entities, spatial entities, environmental entities, and task entities and the fall-in-water risk status entity in the time-series fall-in-water risk knowledge graph, the association weights are adjusted based on the correlation between the contribution of the corresponding key entities in this early warning and the actual risk change results.

[0231] The association weights are updated as follows: for key entities whose contributions are overestimated, the weight of their corresponding associations is reduced by a preset decay coefficient; for key entities whose contributions are underestimated, the weight of their corresponding associations is increased by a preset gain coefficient. The updated association weights are represented as follows: ;

[0232] in, This indicates the association weight before the update. Indicates the learning rate. This represents the weight adjustment amount calculated based on the prediction deviation.

[0233] In step S108, the updated association weights from step S107 are written into the corresponding static association edge attributes in the time-series drowning risk knowledge graph, replacing the original association weight values.

[0234] Simultaneously, the newly added real-time fall-in-water risk facts and their corresponding temporal relationships from the time the warning was issued to the current moment are added to the temporal fall-in-water risk knowledge graph. The newly added real-time fall-in-water risk facts include various changes in personnel status, spatial status, environmental status, and mission status that actually occurred after the warning was issued.

[0235] Based on the updated association weights and newly added risk facts, the risk evolution patterns in the time-series water-fall risk knowledge graph are updated. These patterns include the transition probabilities between various risk states, typical risk evolution path patterns, and the average contribution distribution of each risk factor to changes in risk states.

[0236] When newly added risk facts form a new risk evolution path pattern, the new risk evolution path pattern is added to the historical evolution pattern library of the time-series drowning risk knowledge graph as a candidate pattern for subsequent risk evolution pattern matching.

[0237] Through the above feedback optimization process, the time-series knowledge graph of water-related risks is continuously iterated and updated during operation, and its ability to represent actual risk states and its prediction accuracy are gradually improved.

[0238] Suppose that in a certain warning, the system outputs an interpretable water-fall risk warning of a high-risk level, with a risk probability of 64.27%. The key contributing factors, ranked as follows: wind speed (45.14%), spatial location (platform edge area) (23.70%), and personnel heart rate status (4.51%). Within 10 minutes of the warning being issued, the actual risk status changes as follows: maintenance personnel A did not enter a high-risk state but instead returned to the tower from near the platform edge area, and the risk status decreased from a risk accumulation state to a normal state.

[0239] Comparing the actual changes in risk status with the predicted results, it was found that the risk level prediction bias was that the high risk level was overestimated, while the actual risk level was medium to low. The key causal factor identification bias was that the contribution of wind speed was significantly overestimated, while the contribution of personnel heart rate was underestimated.

[0240] Based on the aforementioned deviations, the associated weights are updated. The association weight between the wind speed entity and the fall-in-water risk entity is reduced by 0.05 according to a preset attenuation coefficient, from 0.45 to 0.40; the association weight between the personnel entity and the fall-in-water risk entity is increased by 0.03 according to a preset gain coefficient, from 0.20 to 0.23. Simultaneously, the newly added risk fact sequence after the warning output—that is, after the wind speed exceeds the limit, the person approaches the platform edge area and then returns to the tower interior before the risk is resolved—is added to the historical evolution pattern library as a new risk evolution pattern.

[0241] After the above updates, the knowledge graph of time-series water-fall risk has improved the accuracy of its predictions for similar risk scenarios. When encountering the initial risk state of approaching the edge of the platform after the wind speed exceeds the standard, it will simultaneously match two possible evolution modes: risk relief after returning to the tower and high risk after attitude imbalance. It will make more accurate predictions based on the specific changing trends of real-time data.

[0242] Reference Figure 2 This application also provides a device for early warning of the risk of offshore wind power maintenance personnel falling into the water, comprising:

[0243] The fusion module 101 is used to form multi-source fused water-fall risk data based on the status information of operation and maintenance personnel, spatial status information, and environmental and task status information.

[0244] The acquisition module 102 is used to acquire a pre-constructed initial knowledge representation of the risk of falling into the water. The initial knowledge representation of the risk of falling into the water is pre-constructed based on historical risk event data and domain prior knowledge, and includes multiple key entities that affect the risk of falling into the water and the static correlation between each key entity.

[0245] The update module 103 is used to update the attributes of the key entity based on the multi-source fused water risk data. When the updated attributes meet the preset risk state triggering conditions, the corresponding attributes are converted into real-time water risk facts with time stamps.

[0246] The module 104 is used to establish a temporal correlation between the real-time water risk facts based on the occurrence time and spatial location of each real-time water risk fact.

[0247] The construction module 105 is used to integrate the key entities, the static relationships, the real-time fall-in-water risk facts and their time information to construct a time-series fall-in-water risk knowledge graph.

[0248] The early warning module 106 is used to reason about the evolution trend of the fall-in-water risk under the combined effect of multiple factors based on the historical evolution pattern in the time-series fall-in-water risk knowledge graph and the real-time fall-in-water risk facts, generate and output interpretable fall-in-water risk early warning information, and the interpretable fall-in-water risk early warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status association path.

[0249] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Those skilled in the art can understand that implementing all or part of the processes of the above embodiments and making equivalent changes according to the claims of this application still fall within the scope of this application.

Claims

1. A method for early warning of the risk of offshore wind power operation and maintenance personnel falling overboard, characterized in that, include: Based on the status information of maintenance personnel, spatial status information, and environmental and task status information, multi-source fusion data on water risk is generated. Obtain a pre-constructed initial knowledge representation of the risk of falling into water. The initial knowledge representation of the risk of falling into water is pre-constructed based on historical risk event data and domain prior knowledge, and includes multiple key entities that affect the risk of falling into water and the static correlation between each key entity. Based on the multi-source fusion water risk data, the attributes of the key entities are updated. When the updated attributes meet the preset risk state triggering conditions, the corresponding attributes are converted into real-time water risk facts with time stamps. Based on the occurrence time and spatial location of each real-time water-fall risk fact, establish the temporal correlation between each real-time water-fall risk fact; By integrating the key entities, the static relationships, the real-time facts of water-falling risk and their time information, a time-series water-falling risk knowledge graph is constructed. Based on the historical evolution patterns in the time-series knowledge graph of falling water risk and the real-time facts of falling water risk, the evolution trend of falling water risk under the combined effect of multiple factors is inferred, and explainable falling water risk early warning information is generated and output. The explainable falling water risk early warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status association paths.

2. The method according to claim 1, characterized in that, The steps of updating the attributes of the key entity based on the multi-source fused fall-in-water risk data, and converting the corresponding attributes into real-time fall-in-water risk facts with time stamps when the updated attributes meet the preset risk state triggering conditions, include: Based on the status information of maintenance personnel in the multi-source fused water risk data, the numerical values ​​of the personnel status attribute set associated with the personnel entity are updated; the personnel status attribute set includes vital sign status attributes and behavioral status attributes, and the time identifier of this update is recorded. Based on the spatial state information in the multi-source fusion water-fall risk data, the spatial risk attribute set associated with the spatial entities is numerically updated, and the time identifier of this update is recorded. Based on the environmental and task status information in the multi-source fusion water risk data, the environmental status attribute set associated with the environmental entity and the task attribute set associated with the task entity are updated numerically, and the time identifier of this update is recorded. According to the preset risk status determination rules, the updated attributes of each key entity are analyzed. When the attribute of any key entity or the relationship between entities meets the preset risk status triggering conditions, the corresponding attribute or relationship is converted into a real-time drowning risk fact with a time stamp.

3. The method according to claim 1, characterized in that, The steps for establishing the temporal correlation between the real-time water-fall risk facts based on their occurrence time and spatial location include: Obtain the entity type, relationship type, occurrence time, and spatial location corresponding to each of the aforementioned real-time water-fall risk facts; Based on the occurrence time t of any two of the aforementioned real-time risk of falling into the water facts i and t j The time correlation coefficient is calculated using the time decay function. : Where k is a preset time decay coefficient, and the time correlation coefficient is... Used to characterize the time dependency in the continuous evolution of risk states; Based on the spatial location p corresponding to any two real-time water risk facts i and p j Calculate the spatial correlation coefficient : The spatial correlation coefficient Used to characterize the spatial constraint relationship between changes in personnel activity areas and hazardous areas; Based on the time correlation coefficient and the spatial correlation coefficient Calculate the overall correlation strength between any two of the aforementioned real-time fall-in-water risk facts. : ； in, The first preset weighting coefficient, The second preset weighting coefficient, This is the offset of the overall association strength for the preset temporal association relationships; When the comprehensive correlation strength When the value exceeds a preset correlation threshold, a temporal correlation is established between the two real-time drowning risk facts, and the overall correlation strength is... As the association weight of the corresponding temporal relationship.

4. The method according to claim 1, characterized in that, Based on the historical evolution patterns in the time-series knowledge graph of falling water risk and the real-time facts of falling water risk, the steps for inferring the evolution trend of falling water risk under the combined influence of multiple factors, generating and outputting interpretable falling water risk early warning information include: Based on the real-time risk facts of falling into the water at the current moment, the corresponding risk evolution pattern is matched in the time-series risk knowledge graph to determine the evolution trend of the current risk status of the maintenance personnel. Based on the evolution trend, the temporal correlation between each key entity in the temporal knowledge graph of drowning risk is obtained, information is disseminated along the temporal correlation path for different risk factors, comprehensive risk correlation characteristics are obtained, and the correlation strength of the current risk state to the target risk state is obtained based on the comprehensive risk correlation characteristics. Based on the correlation strength, the current risk status characteristics are fused with the comprehensive risk correlation characteristics and input into the drowning risk prediction model to calculate the probability that maintenance personnel will enter the target drowning risk state within a preset future time range. Based on the contribution of each key entity to the probability, the key entities affecting the change of the current risk state are determined, the corresponding risk state association paths are extracted, and interpretable drowning risk warning information is generated and output.

5. The method according to claim 4, characterized in that, Based on the real-time risk of falling into the water at the current moment, the steps to match the corresponding risk evolution pattern in the time-series risk knowledge graph and determine the evolution trend of the current risk status of maintenance personnel include: Based on the real-time risk of falling into the water at the current moment, obtain the corresponding risk entity type, entity relationship type, time of occurrence, and risk association weight; Using the risk entity type and the entity association type as search criteria, historical risk facts with the same risk entity type and the same entity association type as the current risk status are extracted from the time-series drowning risk knowledge graph and arranged in chronological order to form a historical risk fact sequence. Calculate the similarity between the temporal features corresponding to the current risk state and the temporal features corresponding to each candidate risk evolution pattern in the historical risk fact sequence; The candidate risk evolution pattern with the highest similarity is identified as the matching risk evolution pattern, and the evolution trend corresponding to the current risk status of the operation and maintenance personnel is determined based on the matching risk evolution pattern.

6. The method according to claim 4, characterized in that, The steps of obtaining the temporal correlation relationships between key entities in the temporal knowledge graph of drowning risk based on the evolutionary trend, propagating information about different risk factors along the temporal correlation path to obtain comprehensive risk correlation characteristics, and obtaining the correlation strength of the evolution from the current risk state to the target risk state based on the comprehensive risk correlation characteristics include: Based on the aforementioned evolutionary trend, the key entities involved in the risk evolution are identified, including personnel entities, space entities, environmental entities, mission entities, and risk status entities. Obtain the temporal association relationships and corresponding temporal association paths among the key entities involved in the risk evolution in the temporal knowledge graph of drowning risk. Along the time-series association path, the risk factor characteristics of each key entity are sequentially transmitted to adjacent entities. When any entity receives risk factor characteristics transmitted by multiple adjacent entities, they are weighted and summed according to the risk association weights corresponding to each association relationship to form the comprehensive risk association characteristics of that entity. Based on the comprehensive risk correlation characteristics, the trend of risk state change under the combined effect of different risk factors is analyzed to obtain the correlation strength of the evolution of the current risk state to the target risk state.

7. The method according to claim 4, characterized in that, The steps of determining the key entities affecting the change of the current risk state based on the contribution of each key entity to the probability, extracting the corresponding risk state association paths, generating and outputting interpretable drowning risk warning information include: The contribution attribution method is used to calculate the degree of contribution of each key entity in the time-series knowledge graph of drowning risk to the probability. The key entities are ranked according to their contribution level, and those whose contribution level exceeds a preset threshold are identified as key entities that affect the change of the current risk status. Based on the key entities, corresponding risk facts, and relationships between entities, the risk state association path from the initial risk state to the current risk state is extracted from the time-series drowning risk knowledge graph. Generate and output interpretable fall-in-water risk warning information, which includes risk level, risk factors and their contribution ranking, risk facts and risk status association path.

8. The method according to claim 1, characterized in that, The method further includes: Based on the deviation between the output of the explainable water risk warning information and the actual risk status change results, the association weights between the key entities and the water risk status entities in the time-series water risk knowledge graph are updated. Based on the updated relationships and newly added real-time water risk facts, the key entities, relationships, and risk evolution patterns in the time-series water risk knowledge graph are updated to optimize the time-series water risk knowledge graph.

9. The method according to claim 1, characterized in that, The key entities include personnel entities, space entities, environmental entities, mission entities, and entities representing the risk of falling into the water. The personnel entity is used to represent personnel objects participating in offshore wind power operation and maintenance activities, and is associated with a personnel status attribute set, which includes vital sign status attributes and behavioral status attributes. The spatial entity is used to characterize the physical space and its spatial risk characteristics involved in offshore wind power operation and maintenance, and to associate a set of spatial risk attributes, which includes a hazard level. The environmental entities are used to characterize external environmental factors that affect the safety of offshore wind power operation and maintenance personnel. The task entity is used to represent the type of work activity currently being performed by the operation and maintenance personnel, and is associated with a task attribute set, which includes the task duration and the task risk level. The water-fall risk status entity is used to characterize the risk evolution status formed by the combined effects of personnel status, environmental factors and task status under specific time and space conditions, including normal status, risk initiation status, risk accumulation status, high-risk status and near-water-fall status.

10. A device for early warning of the risk of offshore wind power maintenance personnel falling overboard, characterized in that, include: The fusion module is used to generate multi-source fused water-fall risk data based on the status information of maintenance personnel, spatial status information, and environmental and task status information. The acquisition module is used to acquire a pre-constructed initial knowledge representation of the risk of falling into the water. The initial knowledge representation of the risk of falling into the water is pre-constructed based on historical risk event data and domain prior knowledge, and includes multiple key entities that affect the risk of falling into the water and the static correlation between each key entity. The update module is used to update the attributes of the key entity based on the multi-source fused water risk data. When the updated attributes meet the preset risk state triggering conditions, the corresponding attributes are converted into real-time water risk facts with time stamps. A module is established to establish a temporal correlation between the real-time water risk facts based on their occurrence time and spatial location. The construction module is used to integrate the key entities, the static relationships, the real-time fall-in-water risk facts and their time information to construct a time-series fall-in-water risk knowledge graph; The early warning module is used to reason about the evolution trend of falling water risk under the combined effect of multiple factors based on the historical evolution pattern in the time-series falling water risk knowledge graph and the real-time falling water risk facts, generate and output interpretable falling water risk early warning information, and the interpretable falling water risk early warning information includes risk level, risk factors and their contribution ranking, risk facts and risk status association path.

Images