Smart factory new energy safety early warning method based on knowledge graph

Abstract

The invention discloses a smart factory new energy safety early warning method based on a knowledge graph, and the method comprises the following steps: collecting multi-source data in the operation of a plurality of devices, constructing a unified time series data set, carrying out the modeling of a complex dependency relationship between various operation parameters through a Vine Copula method, and obtaining the condition dependency intensity between variables; converting the identified abnormal state into an event sequence, establishing a triggering relationship between events by adopting a Hawkes process, reversely adjusting a dependency structure between variables through event triggering strength, realizing dynamic updating of the structure, constructing a risk map structure combining variable dependency and an event triggering mechanism, and carrying out path reasoning to obtain a risk map structure; and identifying the potential risk of the target variable, and outputting early warning information. According to the method, the relationship between data dependence and event influence among equipment can be comprehensively analyzed, and dynamic identification and early warning of potential risks in a complex industrial system are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent monitoring and safety early warning technology, and in particular to a smart factory new energy safety early warning method based on knowledge graph. Background Technology

[0002] With the accelerated development of smart factories, the energy internet, and the digitalization of industrial equipment, the safe operation and management of new energy equipment has become a crucial component of ensuring the reliability of industrial systems. Currently, the large amount of heterogeneous data generated during the operation of new energy equipment is typically collected by sensors and uploaded to edge computing nodes or cloud platforms, where rule matching, feature extraction, or machine learning algorithms are used for status monitoring and anomaly early warning.

[0003] Existing methods for early warning of new energy safety still have significant shortcomings in modeling dependencies between multi-source equipment data and constructing risk propagation mechanisms. On the one hand, traditional methods often use fixed correlation coefficients, distance metrics, or manual feature engineering for variable association modeling, which makes it difficult to reveal the nonlinear, multi-level statistical dependency structure between multi-dimensional operating variables of new energy equipment, resulting in an inability to effectively characterize the true coupling relationship between variables. On the other hand, mainstream anomaly detection algorithms are mostly based on static threshold judgments or classification model outputs, failing to construct dynamic triggering mechanisms between events, unable to reflect complex propagation phenomena such as chain failures and secondary risks, and lacking the ability to link and update the relationship between variable dependencies and event triggering, thus affecting the timeliness and accuracy of early warnings.

[0004] Therefore, how to provide a knowledge graph-based smart factory new energy safety early warning method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based smart factory new energy safety early warning method. This invention utilizes the Vine Copula method to construct a nonlinear dependency structure among multi-source variables of new energy equipment, combines it with Hawkes processes to establish a triggering mechanism between abnormal events, and achieves dynamic coupling and updating of variable dependencies and event propagation intensity. Finally, based on the constructed risk knowledge graph, path reasoning is performed to output risk early warning information for target variables of new energy equipment. This method possesses the ability to model high-dimensional dependencies, characterize chain-like risks, and achieve dynamic evolution and interpretable reasoning.

[0006] A knowledge graph-based smart factory new energy safety early warning method according to an embodiment of the present invention includes the following steps: Collect multi-source heterogeneous operation data from multiple new energy devices in a smart factory, perform preprocessing, and construct a standardized time-series feature set; Based on the operating condition variables in the standardized time series feature set, a high-dimensional dependency structure graph containing multiple nodes and edges is constructed using the Vine Copula method. The high-dimensional dependency structure graph of the operating condition variables of new energy equipment contains multiple operating condition variable nodes and their corresponding condition dependency edge weights. Extract the conditional dependency strength between operating condition variables from the high-dimensional dependency structure graph to form a dependency coefficient matrix; An abnormal event sequence is constructed based on the abnormal states identified in the standardized time series feature set, and the Hawkes process is used to model the abnormal event sequence. The conditional dependency strength between the working condition variables in the dependency coefficient matrix is ​​introduced as the trigger strength adjustment parameter in the Hawkes process, and an abnormal event trigger strength sequence is generated. The high-dimensional dependency structure graph is updated in reverse by using the intensity sequence triggered by the abnormal events generated by the Hawkes process, resulting in a dynamically updated high-dimensional dependency structure graph. A risk knowledge graph containing dynamically updated high-dimensional dependency structure graphs and abnormal event trigger intensity sequences is constructed. Path reasoning is performed based on the risk knowledge graph to output risk warning information for target operating condition variables of new energy equipment.

[0007] Optional preprocessing includes: data cleaning, time alignment, and normalization.

[0008] Optionally, constructing a high-dimensional dependency structure graph containing multiple nodes and edges includes: Collect all operating condition variables of new energy equipment from the standardized time-series feature set, and establish a set of operating condition variables of new energy equipment for data input in the dependency modeling process; For each operating condition variable in the set of new energy equipment operating condition variables, a marginal distribution fitting process is performed to generate a set of marginal distribution functions, and the set of new energy equipment operating condition variables is converted into a set of pseudo-observation data for dependency modeling; Based on the pseudo-observation dataset, the structural type of the Vine Copula model is selected, including the central structure or the chain structure, and the order of the operating condition variables of new energy equipment in the dependency structure modeling is determined. Construct a first-level tree structure. In the first-level tree structure, each node corresponds to a new energy equipment operating condition variable, and each edge connects two operating condition variables and is used to represent their marginal dependencies. Based on the two working condition variables connected by the edges in the first-level tree structure, the set of condition variables is determined, and the second-level tree structure and its subsequent levels are recursively constructed on the basis of the set of condition variables. Each edge in each level represents the conditional dependency relationship between two working condition variables under the given set of condition variables. During the construction of each tree structure, the parameters of the dependency function are estimated, and the two working variables, the level, the set of condition variables and their fitting parameters of each edge are uniformly recorded as structured dependency relationship units. After all tree structure levels are constructed, all working condition variable nodes are extracted as a set of nodes in a high-dimensional dependency structure graph, and all structured dependency relationship units are extracted as an edge set. The high-dimensional dependency structure graph consists of a set of nodes and a set of edges. Each node represents a new energy equipment operating condition variable, and each edge contains two operating condition variables, the tree structure level to which it belongs, the set of condition variables it depends on, and a dependency strength parameter, which are used to represent the multi-level statistical dependency structure between new energy equipment operating condition variables.

[0009] Optionally, the dependency coefficient matrix may include: Extract the complete set of edges from the high-dimensional dependency structure graph, where each edge connects two new energy equipment operating condition variables and contains dependency strength parameters obtained from the modeling process; For each edge in the edge set, extract the two new energy equipment operating condition variables connected by the corresponding edge, determine the unique index position of the two new energy equipment operating condition variables in the new energy equipment operating condition variable set, and record their corresponding dependency strength parameter values. Based on the set of operating condition variables of new energy equipment, a two-dimensional matrix structure is established, in which each row and each column of the two-dimensional matrix corresponds to a new energy equipment operating condition variable. Fill all the new energy equipment operating condition variables and their dependence strength parameters extracted from the edge set into the corresponding positions in the two-dimensional matrix to construct a numerical matrix representing the dependence strength of the variables on the conditions, and generate a dependence coefficient matrix. During the filling process, if two new energy equipment operating condition variables do not form a valid dependency edge in the high-dimensional dependency structure graph, then default values ​​or null values ​​are filled in their corresponding matrix positions to maintain the integrity of the matrix structure.

[0010] Optionally, generating an abnormal event trigger intensity sequence includes: Historical data of all new energy equipment operating condition variables in the standardized time series feature set are monitored, and abnormal states are identified by setting rules. The abnormal states include variable values ​​exceeding the limit, sudden increase in change, or abnormal deviation of dependency relationship. Each identified abnormal state is constructed as an abnormal event unit. Each abnormal event unit includes the abnormal time, abnormal type, the identifier of the operating condition variable of the new energy equipment to which it belongs, and its characteristic value. All abnormal event units are arranged in chronological order to construct a sequence of abnormal events for new energy equipment, which is used to represent all abnormal states observed during actual operation. Extract the dependency strength parameter between the operating condition variable corresponding to the abnormal event unit and other operating condition variables from the dependency coefficient matrix, which is used to construct the set of trigger relationship coefficients between event pairs; The sequence of abnormal events from new energy equipment is used as the input event stream, and the dependency strength parameter is used as the trigger gain input. These are then input into the Hawkes process model to establish the self-excitation and cross-excitation mechanism between abnormal events. The trigger intensity of each abnormal event unit within a given time interval is calculated using the Hawkes process model. Based on the temporal structure and historical occurrence of the abnormal event sequence, the trigger intensity sequence of abnormal events for new energy equipment is output, and the trigger intensity sequence corresponds to the trigger intensity value generated by each event unit. Optionally, the Hawkes procedure includes: Extract each abnormal event unit from the abnormal event sequence of new energy equipment, record its abnormal occurrence time, corresponding new energy equipment operating condition variable identifier and abnormal type, and construct an input event stream arranged in chronological order; Set a base intensity parameter for each type of abnormal event to represent the natural occurrence level of the event under conditions of lack of historical triggers; Extract the conditional dependency strength between the new energy equipment operating condition variables and other operating condition variables corresponding to each abnormal event unit from the dependency coefficient matrix, and form a set of dynamic triggering gain factors associated with the event variables; Based on the temporal position of the abnormal event unit in the abnormal event sequence, the causal transmission path between it and all historical events is identified, the propagation level between abnormal event sequences is determined, and the corresponding chain decay weight set is generated. For each target abnormal event unit, retrieve all historical abnormal events that occurred before its time point, and calculate the triggering contribution value of each historical abnormal event to the target event. When calculating the trigger contribution value of each historical event, the trigger kernel function is called to output the result, and the trigger contribution value is multiplied by the dynamic trigger gain factor and chain decay weight corresponding to the historical abnormal event to obtain the current contribution value. The contribution values ​​of all historical anomalies are summed with the base intensity parameter of the current event to obtain the final trigger intensity value of the corresponding target anomaly event unit. Repeat the above calculation steps for all abnormal event units in the abnormal event sequence, and output the trigger strength values ​​of all abnormal event units in chronological order to form an abnormal event trigger strength sequence.

[0011] Optionally, generating a dynamically updated high-dimensional dependency structure graph includes: Extract the entire set of edges from the high-dimensional dependency structure graph. Each edge connects two new energy equipment operating condition variables and is associated with an original conditional dependency strength. For each edge connecting two new energy equipment operating condition variables, retrieve abnormal event units containing the corresponding two variables in the abnormal event trigger intensity sequence, filter all events occurring within a specified time window, and extract the trigger intensity value of the abnormal event unit. The selected trigger strength values ​​are grouped and calculated according to the working condition variable pairs to obtain the average joint trigger strength of the corresponding edge and variable pair, which is used as the actual activation strength of the variable pair. For each edge, construct a dependency strength correction function based on the numerical relationship between its joint trigger strength mean and the original conditional dependency strength, and output the updated conditional dependency strength. The updated conditional dependency strength is assigned to the corresponding edge structure, replacing the edge parameters in the original edge set, and a new edge set is generated. Keeping the node set in the original high-dimensional dependency structure graph unchanged, the updated edge set is combined with the original node set to generate a dynamically updated high-dimensional dependency structure graph. The dynamically updated high-dimensional dependency structure graph serves as a graphical model representing the structural dependency changes of new energy equipment operating condition variables under the current abnormal state.

[0012] Optionally, the risk warning information for the target operating condition variables of new energy equipment output includes: By integrating the dynamically updated high-dimensional dependency structure graph with the abnormal event trigger intensity sequence, a risk knowledge graph is constructed. The risk knowledge graph includes a set of nodes and a set of directed edges. Nodes represent the operating conditions of new energy equipment and their corresponding abnormal states, while directed edges represent the dependency relationships between variable pairs and the propagation direction of abnormal events. For each edge in the high-dimensional dependency structure graph, the abnormal event unit of the corresponding working condition variable pair in the trigger intensity sequence is matched, its trigger intensity value is extracted, and the condition dependency intensity and trigger intensity are weighted and combined to generate the edge weight of the directed edge, which is used to represent the risk propagation capability between variables. Based on the edge weights of the generated directed edges, abnormal event nodes with trigger strength greater than a set trigger threshold are marked in the risk knowledge graph and used as initial risk source nodes. Starting from each initial risk source node, traverse all reachable paths according to the risk knowledge graph structure, record all nodes passed through on the path, edge weights and propagation order, and calculate the cumulative risk propagation intensity of each path; The path set generated by the traversal is established as a path record set. Each path record includes the path start point, path end point, sequence of operating condition variables of new energy equipment passed through, cumulative risk propagation intensity, and path depth. The path in the selected path record set that has the target new energy equipment operating condition variable as the path endpoint and whose cumulative risk propagation intensity is higher than the set risk propagation threshold is selected as the candidate risk path. The path structure, target operating condition variable identifier, cumulative risk propagation intensity, and corresponding trigger time window in the candidate risk paths are compiled into risk warning information for the target operating condition variables of new energy equipment.

[0013] The beneficial effects of this invention are: (1) This invention introduces the Vine Copula method to construct a high-dimensional nonlinear dependency structure between the operating conditions of new energy equipment, which can accurately characterize the multi-level correlation between multi-source heterogeneous data, break through the applicability limitations of traditional linear modeling and feature engineering in complex industrial scenarios, and enable the coupling mode between variables to be represented in a structured manner, providing a structural basis for subsequent risk propagation modeling.

[0014] (2) This invention establishes a self-excitation and cross-excitation triggering mechanism between abnormal events by introducing the Hawkes process, and further proposes a two-way dynamic coupling method between dependency structure and event intensity, which realizes the real-time feedback update of the variable dependency structure by the abnormal event triggering mechanism, and enhances the adaptability and expressiveness of the model to the risk evolution process.

[0015] (3) This invention constructs a risk knowledge graph that integrates operating condition variables, abnormal events, dependencies and triggering paths, supports path reasoning based on graph structure, can identify the risk propagation chain of target operating condition variables starting from the risk source node, improve the traceability and interpretability of early warning results, and realize forward perception and accurate early warning of complex chain failures and secondary risks. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a knowledge graph-based smart factory new energy safety early warning method proposed in this invention; Figure 2 This is a flowchart illustrating the modeling process for constructing anomaly event trigger intensity sequences based on Hawkes procedures in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-2 A knowledge graph-based method for early warning of new energy safety in smart factories includes the following steps: Collect multi-source heterogeneous operation data from multiple new energy devices in a smart factory, perform preprocessing, and construct a standardized time-series feature set; Based on the operating condition variables in the standardized time series feature set, a high-dimensional dependency structure graph containing multiple nodes and edges is constructed using the Vine Copula method. The high-dimensional dependency structure graph of the operating condition variables of new energy equipment contains multiple operating condition variable nodes and their corresponding condition dependency edge weights. Extract the conditional dependency strength between operating condition variables from the high-dimensional dependency structure graph to form a dependency coefficient matrix; An abnormal event sequence is constructed based on the abnormal states identified in the standardized time series feature set, and the Hawkes process is used to model the abnormal event sequence. The conditional dependency strength between the working condition variables in the dependency coefficient matrix is ​​introduced as the trigger strength adjustment parameter in the Hawkes process, and an abnormal event trigger strength sequence is generated. The high-dimensional dependency structure graph is updated in reverse by using the intensity sequence triggered by the abnormal events generated by the Hawkes process, resulting in a dynamically updated high-dimensional dependency structure graph. A risk knowledge graph containing dynamically updated high-dimensional dependency structure graphs and abnormal event trigger intensity sequences is constructed. Path reasoning is performed based on the risk knowledge graph to output risk warning information for target operating condition variables of new energy equipment.

[0019] In this embodiment, preprocessing includes: data cleaning, time alignment, and normalization.

[0020] In this embodiment, constructing a high-dimensional dependency structure graph containing multiple nodes and edges includes: Collect all operating condition variables of new energy equipment from the standardized time-series feature set, and establish a set of operating condition variables of new energy equipment for data input in the dependency modeling process; For each operating condition variable in the set of new energy equipment operating condition variables, a marginal distribution fitting process is performed to generate a set of marginal distribution functions, and the set of new energy equipment operating condition variables is converted into a set of pseudo-observation data for dependency modeling; Marginal distribution fitting refers to selecting an appropriate probability distribution model to characterize the marginal distribution characteristics of each new energy equipment's operating condition variable based on its historical observation samples. Parametric methods such as normal distribution, gamma distribution, or log-normal distribution can be used for fitting, or non-parametric methods such as empirical distribution functions or kernel density estimation can be used for estimation. The selection of the distribution model is determined based on the goodness-of-fit test results. Commonly used test methods include the Kolmogorov-Smirnov test or the Akaike information criterion. The marginal distribution function obtained through fitting is used to map the original variables to the [0,1] interval. Based on the pseudo-observation dataset, the structural type of the Vine Copula model is selected, including the central structure or the chain structure, and the order of the operating condition variables of new energy equipment in the dependency structure modeling is determined. Construct a first-level tree structure. In the first-level tree structure, each node corresponds to a new energy equipment operating condition variable, and each edge connects two operating condition variables and is used to represent their marginal dependencies. Based on the two working condition variables connected by the edges in the first-level tree structure, the set of condition variables is determined, and the second-level tree structure and its subsequent levels are recursively constructed on the basis of the set of condition variables. Each edge in each level represents the conditional dependency relationship between two working condition variables under the given set of condition variables. During the construction of each tree structure, the parameters of the dependency function are estimated, and the two working variables, the level, the set of condition variables and their fitting parameters of each edge are uniformly recorded as structured dependency relationship units. Dependency function parameter estimation refers to the numerical solution of the parameters of the Copula function type selected for each pair of working condition variables during Vine Copula modeling. The commonly used method is maximum likelihood estimation, which involves constructing the log-likelihood function of the corresponding Copula function based on known pseudo-observation data, and solving for the optimal parameter values ​​through numerical optimization algorithms so that the selected Copula function fits the dependency structure between variables to the greatest extent. If the Copula function has multiple candidate forms, the model can be optimized through the Akaike information criterion or the Bayesian information criterion. The parameter estimation results are used to determine the dependency strength of the edges and construct the final high-dimensional dependency structure graph. The fitting parameters are derived from modeling the joint distribution between variable pairs in pseudo-observation data. By selecting an appropriate Copula function form and solving for its parameters using the maximum likelihood estimation method, the optimal parameters obtained are the fitting parameters, which are used to quantify the dependence strength between variables. After all tree structure levels are constructed, all working condition variable nodes are extracted as a set of nodes in a high-dimensional dependency structure graph, and all structured dependency relationship units are extracted as an edge set. The high-dimensional dependency structure graph consists of a set of nodes and a set of edges. Each node represents a new energy equipment operating condition variable, and each edge contains two operating condition variables, the tree structure level to which it belongs, the set of condition variables it depends on, and the dependency strength parameter, which are used to represent the multi-level statistical dependency structure between the new energy equipment operating condition variables. The dependency strength parameter is derived from the optimal parameter value obtained by fitting the selected Copula function during the dependency function modeling process between each pair of new energy equipment operating condition variables. It is used to quantitatively describe the strength of the dependency relationship between the variable pair under a specific level and set of conditions, and is one of the core attributes of the edges in the high-dimensional dependency structure graph.

[0021] In this embodiment, the dependency coefficient matrix includes: Extract the complete set of edges from the high-dimensional dependency structure graph, where each edge connects two new energy equipment operating condition variables and contains dependency strength parameters obtained from the modeling process; For each edge in the edge set, extract the two new energy equipment operating condition variables connected by the corresponding edge, determine the unique index position of the two new energy equipment operating condition variables in the new energy equipment operating condition variable set, and record their corresponding dependency strength parameter values. Based on the set of operating condition variables of new energy equipment, a two-dimensional matrix structure is established, in which each row and each column of the two-dimensional matrix corresponds to a new energy equipment operating condition variable. Fill all the new energy equipment operating condition variables and their dependence strength parameters extracted from the edge set into the corresponding positions in the two-dimensional matrix to construct a numerical matrix representing the dependence strength of the variables on the conditions, and generate a dependence coefficient matrix. During the filling process, if two new energy equipment operating condition variables do not form a valid dependency edge in the high-dimensional dependency structure graph, then default values ​​or null values ​​are filled in their corresponding matrix positions to maintain the integrity of the matrix structure.

[0022] In this embodiment, generating the abnormal event trigger strength sequence includes: Historical data of all new energy equipment operating condition variables in the standardized time series feature set are monitored, and abnormal states are identified by setting rules. The abnormal states include variable values ​​exceeding the limit, sudden increase in change, or abnormal deviation of dependency relationship. The rules set refer to the criteria used to identify abnormal states of operating variables of new energy equipment. Specifically, they include the following categories: First, static threshold rules, which mark an abnormality when the operating variable exceeds its design upper or lower limit; second, sliding statistical rules, such as the mean and variance calculated based on a sliding window, which trigger an abnormality when the variable deviates from the normal range by more than a set multiple; third, rate of change rules, which detect whether the increase or decrease of a variable in a unit of time exceeds a preset rate threshold; and fourth, correlation deviation rules, which mark a collaborative anomaly when there is a significant deviation between the actual dependence strength between variables and the expected strength in the high-dimensional dependence structure diagram. Each identified abnormal state is constructed as an abnormal event unit. Each abnormal event unit includes the abnormal time, abnormal type, the identifier of the operating condition variable of the new energy equipment to which it belongs, and its characteristic value. All abnormal event units are arranged in chronological order to construct a sequence of abnormal events for new energy equipment, which is used to represent all abnormal states observed during actual operation. Extract the dependency strength parameter between the operating condition variable corresponding to the abnormal event unit and other operating condition variables from the dependency coefficient matrix, which is used to construct the set of trigger relationship coefficients between event pairs; The sequence of abnormal events from new energy equipment is used as the input event stream, and the dependency strength parameter is used as the trigger gain input. These are then input into the Hawkes process model to establish the self-excitation and cross-excitation mechanism between abnormal events. The trigger intensity of each abnormal event unit within a given time interval is calculated using the Hawkes process model. Based on the temporal structure and historical occurrence of the abnormal event sequence, the trigger intensity sequence of abnormal events for new energy equipment is output, and the trigger intensity sequence corresponds to the trigger intensity value generated by each event unit. Based on the temporal structure and historical occurrence of anomalous event sequences, the modeling process comprehensively considers information such as the order of anomalous events on the timeline, their frequency and interval distribution, as well as the triggering relationships and persistence patterns among previously observed anomalous events, to reflect the dynamic evolution trend of the events. This information serves as the input basis for the Hawkes process model, helping to derive the trigger probability of the current event and further generate trigger intensity values ​​reflecting self-excitation and cross-excitation effects.

[0023] In this embodiment, the Hawkes process includes: Extract each abnormal event unit from the abnormal event sequence of new energy equipment, record its abnormal occurrence time, corresponding new energy equipment operating condition variable identifier and abnormal type, and construct an input event stream arranged in chronological order; Set a base intensity parameter for each type of abnormal event to represent the natural occurrence level of the event under conditions of lack of historical triggers; Extract the conditional dependency strength between the new energy equipment operating condition variables and other operating condition variables corresponding to each abnormal event unit from the dependency coefficient matrix, and form a set of dynamic triggering gain factors associated with the event variables; Based on the temporal position of the abnormal event unit in the abnormal event sequence, the causal transmission path between it and all historical events is identified, the propagation level between abnormal event sequences is determined, and the corresponding chain decay weight set is generated. Identifying the causal transmission path between the event and all historical events means that, for the current abnormal event unit, according to the time order in the abnormal event sequence, it is determined whether there is a potential triggering relationship between all historical abnormal events that are earlier than the current event. The criteria for this determination include: whether the operating condition variables of the new energy equipment corresponding to the two events have a non-zero dependency strength in the dependency coefficient matrix, and whether the time interval between the occurrence of the two events is within the set triggering time window. Historical events that meet the above conditions are considered to have a causal impact on the current event and form a transmission path according to the time hierarchy. For each target abnormal event unit, retrieve all historical abnormal events that occurred before its time point, and calculate the triggering contribution value of each historical abnormal event to the target event. When calculating the trigger contribution value of each historical event, the trigger kernel function is called to output the result, and the trigger contribution value is multiplied by the dynamic trigger gain factor and chain decay weight corresponding to the historical abnormal event to obtain the current contribution value. The trigger kernel function is a function used to characterize the decreasing trend of the influence of historical anomalous events on future anomalous events over time. After inputting the occurrence time of the historical anomalous event and the current event time, the trigger kernel function outputs a non-negative value based on the time interval between the two. This value represents the trigger contribution of the historical event to the current event. The trigger kernel function usually has the characteristic of monotonically decreasing as the time interval increases. This can be achieved through exponential decay or other calculable decay forms. Its output value participates in the calculation of trigger strength in the subsequent calculation along with the dynamic trigger gain factor and the chain decay weight. The contribution values ​​of all historical anomalies are summed with the base intensity parameter of the current event to obtain the final trigger intensity value of the corresponding target anomaly event unit. Repeat the above calculation steps for all abnormal event units in the abnormal event sequence, and output the trigger strength values ​​of all abnormal event units in chronological order to form an abnormal event trigger strength sequence.

[0024] In this embodiment, generating the dynamically updated high-dimensional dependency structure graph includes: Extract the entire set of edges from the high-dimensional dependency structure graph. Each edge connects two new energy equipment operating condition variables and is associated with an original conditional dependency strength. For each edge connecting two new energy equipment operating condition variables, retrieve abnormal event units containing the corresponding two variables in the abnormal event trigger intensity sequence, filter all events occurring within a specified time window, and extract the trigger intensity value of the abnormal event unit. The selected trigger strength values ​​are grouped and calculated according to the working condition variable pairs to obtain the average joint trigger strength of the corresponding edge and variable pair, which is used as the actual activation strength of the variable pair. For each edge, construct a dependency strength correction function based on the numerical relationship between its joint trigger strength mean and the original conditional dependency strength, and output the updated conditional dependency strength. The dependency strength correction function is a mechanism that adjusts the dependency parameters of edges in a high-dimensional dependency structure graph based on the relationship between the trigger strength of abnormal events in new energy equipment and the original conditional dependency strength. Specifically, when the joint trigger strength of a variable pair in the trigger strength sequence is significantly higher than the historical dependency value, the dependency strength of that edge is appropriately increased; conversely, it is decreased. This function is typically implemented using linear weighting, difference mapping, or piecewise adjustment to ensure that the updated dependency parameters change continuously within a reasonable range and can dynamically reflect substantial changes in the relationship between variables as events evolve. The updated conditional dependency strength is assigned to the corresponding edge structure, replacing the edge parameters in the original edge set, and a new edge set is generated. In a high-dimensional dependency structure graph, the corresponding edge structure refers to the data unit used to represent the conditional dependency relationship between two new energy equipment operating condition variables. Each edge structure contains the identifier of the variable pair, its modeling level, the set of condition variables, and the corresponding dependency strength parameter. When dynamically updated, the dependency strength parameter in the edge structure will be replaced or adjusted according to the strength feedback triggered by abnormal events to reflect the change in the statistical association relationship between the current variable pair, while other structural information of the edge remains unchanged. Keeping the node set in the original high-dimensional dependency structure graph unchanged, the updated edge set is combined with the original node set to generate a dynamically updated high-dimensional dependency structure graph. The dynamically updated high-dimensional dependency structure graph serves as a graphical model representing the structural dependency changes of new energy equipment operating condition variables under the current abnormal state.

[0025] In this embodiment, the risk warning information for outputting the target operating condition variables of new energy equipment includes: By integrating the dynamically updated high-dimensional dependency structure graph with the abnormal event trigger intensity sequence, a risk knowledge graph is constructed. The risk knowledge graph includes a set of nodes and a set of directed edges. Nodes represent the operating conditions of new energy equipment and their corresponding abnormal states, while directed edges represent the dependency relationships between variable pairs and the propagation direction of abnormal events. For each edge in the high-dimensional dependency structure graph, the abnormal event unit of the corresponding working condition variable pair in the trigger intensity sequence is matched, its trigger intensity value is extracted, and the condition dependency intensity and trigger intensity are weighted and combined to generate the edge weight of the directed edge, which is used to represent the risk propagation capability between variables. Weighted combination refers to the numerical fusion of the conditional dependency strength of variable pairs in a high-dimensional dependency structure graph with the corresponding abnormal event triggering strength when constructing a risk knowledge graph. This fusion is used to calculate the final weight value of the edges in the risk propagation path. Specifically, the original conditional dependency strength of a certain working condition variable pair and its corresponding joint triggering strength in the abnormal event triggering strength sequence are extracted and combined according to a set proportional coefficient or linear weighting function to obtain the edge weight that reflects the risk propagation capability of the variable pair in the current state. This weight is used for path propagation and cumulative risk value calculation in the time series graph. Based on the edge weights of the generated directed edges, abnormal event nodes with trigger strength greater than a set trigger threshold are marked in the risk knowledge graph and used as initial risk source nodes. Setting a trigger threshold refers to pre-setting a fixed or dynamically calculated numerical limit in the trigger intensity sequence of abnormal events in new energy equipment. When the trigger intensity value of a certain abnormal event unit is greater than the threshold, it is considered to have the potential risk propagation capability. Starting from each initial risk source node, traverse all reachable paths according to the risk knowledge graph structure, record all nodes passed through on the path, edge weights and propagation order, and calculate the cumulative risk propagation intensity of each path; The path set generated by the traversal is established as a path record set. Each path record includes the path start point, path end point, sequence of operating condition variables of new energy equipment passed through, cumulative risk propagation intensity, and path depth. The path in the selected path record set that has the target new energy equipment operating condition variable as the path endpoint and whose cumulative risk propagation intensity is higher than the set risk propagation threshold is selected as the candidate risk path. Setting a risk propagation threshold refers to the judgment boundary set for the cumulative risk value in the risk knowledge graph. When the cumulative risk intensity of a certain propagation path is greater than the threshold, the path is considered to have an actual risk impact effect. This is used to screen the early warning output path of the target new energy equipment operating condition variables. The path structure, target operating condition variable identifier, cumulative risk propagation intensity, and corresponding trigger time window in the candidate risk paths are compiled into risk warning information for the target operating condition variables of new energy equipment.

[0026] Example 1

[0027] To verify the feasibility of this invention in practice, it was applied to the operation and management scenario of new energy equipment in a smart factory. This included the collaborative operation of multiple types of equipment such as wind power converters, battery management units (BMS), energy storage inverters, and high-voltage DC converters. The system as a whole exhibits complex equipment types, high variable dimensionality, strong nonlinear relationships between variables, and dynamic evolution with changing operating conditions. Traditional methods struggle to characterize the nonlinear coupling structure between equipment variables and the time-series triggering mechanisms of abnormal events, resulting in the inability to achieve accurate and forward-looking safety warnings.

[0028] To address the aforementioned issues, this invention provides a knowledge graph-based early warning method for new energy safety. In practical applications, multi-dimensional operational data, including temperature, current, voltage, power, frequency, and operating status codes, are first collected from multiple devices. This data is then uniformly processed using missing data completion and scale normalization to construct a standardized time-series feature set. This set serves as the foundation for subsequent modeling input.

[0029] Based on a standardized time-series feature set, the Vine Copula method is used to model high-order nonlinear dependencies among multiple variables. First, marginal distribution fitting is performed for each variable, with Beta, Gaussian, or t-distribution commonly used for optimal fitting. The Copula tree structure type is determined by a structure selection algorithm, and variable pairs are connected in the first layer to represent marginal dependencies. In each tree layer, the connected variable pairs from the previous layer are used as condition variables, and the edge functions are recursively constructed and the parameters are estimated, finally obtaining a high-dimensional dependency structure graph containing all variables.

[0030] This high-dimensional dependency structure graph consists of a set of nodes and a set of edges. Nodes represent working condition variables, and edges represent variable pairs with conditional dependencies. Each edge is accompanied by attributes such as edge position, Copula function type, and dependency strength, which are used for subsequent event propagation modeling and graph structure updates.

[0031] During time-series monitoring, this invention uses set rules (such as a fixed window change rate greater than 3σ) to identify abnormal variable states and generate abnormal event units, which include an abnormal timestamp, variable ID, abnormal type, and abnormal value. The abnormal events are arranged in chronological order to form an abnormal event sequence, which, along with the aforementioned dependency structure graph, is input into the Hawkes process model.

[0032] The Hawkes process uses the base trigger strength of each event as the underlying signal and introduces a dependency strength as the trigger gain to construct a propagation mechanism between event pairs. At each event time point, the system retrieves the historical event trigger chain, calls the excitation kernel function (e.g., exponential kernel) for each historical event, outputs the decay weight, and calculates the final trigger contribution by combining it with the gain factor. All historical contribution values ​​are accumulated and added to the base strength to obtain the total trigger strength value of the target event.

[0033] Using the above method, a complete sequence of abnormal event trigger strengths is obtained. The system further corrects the original dependency structure graph based on these trigger strength values, comparing the joint strength of the variable pairs involved in the event with the initial edge parameters, and updating it using a correction function to obtain a dynamically updated high-dimensional dependency structure graph.

[0034] By combining the updated structure graph and the abnormal event trigger sequence, a time-series risk knowledge graph is constructed. Each event is treated as a graph node, and the dependent propagation path is treated as a directed edge. Each edge is accompanied by a time decay factor, gain coefficient, and propagation probability. The system outputs the risk value of the target variable through path traversal and risk aggregation calculation. The following are some of the structure data and system output results in this embodiment: Table 1. Strength of conditional dependence between variables (modeling stage) Operating condition variables Conditional Dependency Strength Battery temperature and discharge current 0.81 Discharge current and voltage 0.73 Voltage and power factor 0.67 Wind power converter output voltage and system frequency 0.62 Energy storage load current and high voltage frequency 0.58 Table 2. Trigger strength and propagation delay between anomalous event pairs (Hawkes modeling) Startup event variable Target event variable Trigger strength Propagation delay (minutes) Abnormal battery temperature Discharge current fluctuation 0.42 4 Discharge current fluctuation abnormal output voltage of wind power converter 0.31 5 Voltage abnormality Energy storage module frequency offset 0.27 3 Table 3 Example of Early Warning Result Output Target variable Early warning path (node ​​chain) Risk warning lead time (minutes) Energy storage frequency fluctuation Battery temperature → current → voltage → system frequency → energy storage frequency 11 abnormal output voltage of wind power converter Battery temperature → Voltage → Wind power voltage 8 A comprehensive analysis of the data in Tables 1 to 3 reveals that the method of this invention possesses advantages in structural interpretability and prediction accuracy when applied to complex systems of new energy equipment. Table 1 shows strong conditional dependencies among multiple key variable pairs; for example, the "battery temperature-discharge current" relationship reaches 0.81, indicating that this invention effectively exploits the nonlinear coupling characteristics between multi-source heterogeneous variables using the Vine Copula method, providing a structural foundation for subsequent anomaly propagation paths. Table 2 shows that the trigger intensity distribution and dependency relationships between anomaly events match well, and the event propagation time exhibits a chain-like propagation pattern. For example, the trigger intensity of "discharge current fluctuation" caused by "abnormal battery temperature" is 0.42, taking only 4 minutes, reflecting that the Hawkes process accurately models the self-excitation effect and propagation sequence between events. Table 3 shows that the system possesses a clear prediction chain when identifying high-risk variable paths. For example, the path "battery temperature → current → voltage → frequency → energy storage frequency" can output early warning results within 11 minutes, with a complete path and clear nodes, demonstrating causal reasoning capabilities supported by a dynamic knowledge graph. In summary, the structural data in the three tables fully demonstrates the collaborative closed loop of this invention from high-dimensional dependency structure modeling, event sequence triggering identification to path warning output. It has traceability, dynamic adaptability and structural interpretability, verifying the engineering practicality and technological breakthrough of the method.

[0035] The implementation results demonstrate that this invention effectively supports risk prediction and path reasoning for new energy equipment by describing nonlinear variable relationships using Vine Copula, modeling abnormal event trigger chains using Hawkes processes, and constructing a dynamically evolving knowledge graph. This method not only possesses timeliness and structural interpretability but also supports dynamic feedback updates, improving the response efficiency and identification accuracy to system-level anomaly propagation.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based smart factory new energy safety early warning method, characterized in that, include: Collect multi-source heterogeneous operation data from multiple new energy devices in a smart factory, perform preprocessing, and construct a standardized time-series feature set; Based on the working condition variables in the standardized time series feature set, a high-dimensional dependency structure graph containing multiple nodes and edges is constructed using the Vine Copula method. Extract the conditional dependency strength between operating condition variables from the high-dimensional dependency structure graph to form a dependency coefficient matrix; An abnormal event sequence is constructed based on the abnormal states identified in the standardized time series feature set, and the Hawkes process is used to model the abnormal event sequence. The conditional dependency strength between the working condition variables in the dependency coefficient matrix is ​​introduced as the trigger strength adjustment parameter in the Hawkes process, and an abnormal event trigger strength sequence is generated. The high-dimensional dependency structure graph is updated in reverse by using the intensity sequence triggered by the abnormal events generated by the Hawkes process, resulting in a dynamically updated high-dimensional dependency structure graph. A risk knowledge graph containing dynamically updated high-dimensional dependency structure graphs and abnormal event trigger intensity sequences is constructed. Path reasoning is performed based on the risk knowledge graph to output risk warning information for target operating condition variables of new energy equipment.

2. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, Preprocessing includes: Data cleaning, time alignment, and normalization.

3. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, Constructing a high-dimensional dependency structure graph containing multiple nodes and edges includes: Collect all operating condition variables of new energy equipment from the standardized time-series feature set, and establish a set of operating condition variables of new energy equipment; For each operating condition variable in the set of new energy equipment operating condition variables, a marginal distribution fitting process is performed to generate a set of marginal distribution functions, and the set of new energy equipment operating condition variables is converted into a set of pseudo-observation data for dependency modeling; Based on the pseudo-observation dataset, the structure type of the Vine Copula model is selected, and the order of the operating condition variables of new energy equipment in the dependency structure modeling is determined. Construct a first-level tree structure. In the first-level tree structure, each node corresponds to a new energy equipment operating condition variable, and each edge connects two operating condition variables and is used to represent their marginal dependencies. Based on the two working condition variables connected by the edges in the first-level tree structure, the set of condition variables is determined, and the second-level tree structure and its subsequent levels are recursively constructed on the basis of the set of condition variables. Each edge in each level represents the conditional dependency relationship between two working condition variables under the given set of condition variables. During the construction of each tree structure, the parameters of the dependency function are estimated, and the two working variables, the level, the set of condition variables and their fitting parameters of each edge are uniformly recorded as structured dependency relationship units. After all tree structure levels are constructed, all working condition variable nodes are extracted as a set of nodes in a high-dimensional dependency structure graph, and all structured dependency relationship units are extracted as an edge set. The high-dimensional dependency structure graph consists of a set of nodes and a set of edges. Each node represents a new energy equipment operating condition variable, and each edge contains two operating condition variables, the tree structure level to which it belongs, the set of condition variables it depends on, and a dependency strength parameter, which are used to represent the multi-level statistical dependency structure between new energy equipment operating condition variables.

4. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, The dependency coefficient matrix consists of: Extract the complete set of edges from the high-dimensional dependency structure graph, where each edge connects two new energy equipment operating condition variables and contains dependency strength parameters obtained from the modeling process; For each edge in the edge set, extract the two new energy equipment operating condition variables connected by the corresponding edge, determine the unique index position of the two new energy equipment operating condition variables in the new energy equipment operating condition variable set, and record their corresponding dependency strength parameter values. Based on the set of operating condition variables of new energy equipment, a two-dimensional matrix structure is established, in which each row and each column of the two-dimensional matrix corresponds to a new energy equipment operating condition variable. All the operating condition variables of new energy equipment extracted from the edge set and their dependence strength parameters are filled into the corresponding positions in the two-dimensional matrix to construct a numerical matrix representing the dependence strength of the variables on the conditions, and generate a dependence coefficient matrix.

5. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, The sequence of events to be generated includes: Historical data of all new energy equipment operating condition variables in the standardized time series feature set are monitored, and abnormal states are identified by setting rules; Each identified abnormal state is constructed as an abnormal event unit. Each abnormal event unit includes the abnormal time, abnormal type, the identifier of the operating condition variable of the new energy equipment to which it belongs, and its characteristic value. Arrange all abnormal event units in chronological order to construct a sequence of abnormal events for new energy equipment. Extract the dependency strength parameter between the operating condition variable corresponding to the abnormal event unit and other operating condition variables from the dependency coefficient matrix; The sequence of abnormal events from new energy equipment is used as the input event stream, and the dependency strength parameter is used as the trigger gain input. These are then input into the Hawkes process model to establish the self-excitation and cross-excitation mechanism between abnormal events. The trigger intensity of each abnormal event unit within a given time interval is calculated using the Hawkes process model. Based on the temporal structure and historical occurrence of the abnormal event sequence, a trigger intensity sequence of abnormal events for new energy equipment is output, and the trigger intensity sequence corresponds to the trigger intensity value generated by each event unit.

6. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, The Hawkes process includes: Extract each abnormal event unit from the abnormal event sequence of new energy equipment, record its abnormal occurrence time, corresponding new energy equipment operating condition variable identifier and abnormal type, and construct an input event stream arranged in chronological order; Set the basic strength parameters for each type of abnormal event; Extract the conditional dependency strength between the new energy equipment operating condition variables and other operating condition variables corresponding to each abnormal event unit from the dependency coefficient matrix to form a set of dynamic triggering gain factors; Based on the temporal position of the abnormal event unit in the abnormal event sequence, the causal transmission path between it and all historical events is identified, the propagation level between abnormal event sequences is determined, and the corresponding chain decay weight set is generated. For each target abnormal event unit, retrieve all historical abnormal events that occurred before its time point, and calculate the triggering contribution value of each historical abnormal event to the target event. When calculating the trigger contribution value of each historical event, the trigger kernel function is called to output the result, and the trigger contribution value is multiplied by the dynamic trigger gain factor and chain decay weight corresponding to the historical abnormal event to obtain the current contribution value. The contribution values ​​of all historical anomalies are summed with the base intensity parameter of the current event to obtain the final trigger intensity value of the corresponding target anomaly event unit. Repeat the above calculation steps for all abnormal event units in the abnormal event sequence, and output the trigger strength values ​​of all abnormal event units in chronological order to form an abnormal event trigger strength sequence.

7. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, Generating a dynamically updated high-dimensional dependency structure graph includes: Extract the entire set of edges from the high-dimensional dependency structure graph. Each edge connects two new energy equipment operating condition variables and is associated with an original conditional dependency strength. For each edge connecting two new energy equipment operating condition variables, retrieve abnormal event units containing the corresponding two variables in the abnormal event trigger intensity sequence, filter all events occurring within a specified time window, and extract the trigger intensity value of the abnormal event unit. The selected trigger strength values ​​are grouped and calculated according to the working condition variable pairs to obtain the joint trigger strength mean of the corresponding edge and corresponding variable pairs; For each edge, construct a dependency strength correction function based on the numerical relationship between its joint trigger strength mean and the original conditional dependency strength, and output the updated conditional dependency strength. The updated conditional dependency strength is assigned to the corresponding edge structure, replacing the edge parameters in the original edge set, and a new edge set is generated. Keeping the node set in the original high-dimensional dependency graph unchanged, the updated edge set is combined with the original node set to generate a dynamically updated high-dimensional dependency graph.

8. The method for early warning of new energy safety in smart factories based on knowledge graphs according to claim 1, characterized in that, Risk warning information for outputting target operating condition variables of new energy equipment includes: By integrating the dynamically updated high-dimensional dependency structure graph with the abnormal event trigger intensity sequence, a risk knowledge graph is constructed. The risk knowledge graph includes a set of nodes and a set of directed edges. Nodes represent the operating conditions of new energy equipment and their corresponding abnormal states, while directed edges represent the dependency relationships between variable pairs and the propagation direction of abnormal events. For each edge in the high-dimensional dependency structure graph, match the abnormal event unit of the corresponding working condition variable pair in the trigger intensity sequence, extract its trigger intensity value, and weight the condition dependency intensity and the trigger intensity to generate the edge weight of the directed edge. Based on the edge weights of the generated directed edges, abnormal event nodes with trigger strength greater than a set trigger threshold are marked in the risk knowledge graph and used as initial risk source nodes. Starting from each initial risk source node, traverse all reachable paths according to the risk knowledge graph structure, record all nodes passed through on the path, edge weights and propagation order, and calculate the cumulative risk propagation intensity of each path; The path set generated by the traversal is established as a path record set. Each path record includes the path start point, path end point, sequence of operating condition variables of new energy equipment passed through, cumulative risk propagation intensity, and path depth. The path in the selected path record set that has the target new energy equipment operating condition variable as the path endpoint and whose cumulative risk propagation intensity is higher than the set risk propagation threshold is selected as the candidate risk path. The path structure, target operating condition variable identifier, cumulative risk propagation intensity, and corresponding trigger time window in the candidate risk paths are compiled into risk warning information for the target operating condition variables of new energy equipment.

Images