Multimodal-based power safety protection equipment risk tracing system

Abstract

The application relates to the technical field of power safety and multi-modal data analysis, in particular to a power safety protection equipment risk tracing system based on multi-modal data, which comprises a data alignment unit, an event causal graph construction unit, a causal chain generation unit, a fuzziness quantification unit and a convergence judgment unit.

Description

Multimodal-based risk tracing system for power safety protection equipment Technical Field

[0001] This invention relates to the fields of power safety and multimodal data analysis technology, specifically a risk tracing system for power safety protection equipment based on multimodal data analysis. Background Technology

[0002] In the field of power safety protection, when a safety incident such as a substation tripping or a line fault occurs, the cause of the incident must be traced.

[0003] This tracing process heavily relies on data such as video streams, sensor logs, and operation records. However, this data is multimodal, heterogeneous, and asynchronous in time, and is often accompanied by the loss of key evidence. Existing technologies struggle to effectively integrate such complex data, leading to difficulties in tracing the cause and high ambiguity in the constructed causal chains. Specifically, the system cannot effectively distinguish between multiple possible competing causal chains, nor can it automatically identify the key evidence points that should be prioritized for verification when evidence is insufficient or the system is confused, resulting in interruptions or inefficiencies in the tracing work. Therefore, how to solve the problems of tracing difficulties and high ambiguity in causal chains caused by multimodal data, asynchronous time, and lost evidence has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a multimodal power safety protection equipment risk tracing system. Specifically, the technical solution of this invention includes:

[0005] A multimodal power safety protection equipment risk tracing system includes:

[0006] The data alignment unit is used to collect raw multimodal data and perform normalization processing to generate initial evidence nodes, and then construct an event causal graph.

[0007] The causal chain generation unit is used to search for competing causal chains based on the event causal graph and to calculate the chain likelihood score.

[0008] Fuzzy quantification unit, used to calculate normalized probability based on chain likelihood score and calculate inter-chain fuzziness score;

[0009] The convergence determination unit is used to compare the inter-chain ambiguity score with the preset narrative convergence threshold to obtain the convergence signal or the signal to be verified.

[0010] When a signal to be verified is generated, the evidence identification unit is used to calculate the estimated information gain and sort it to obtain the key evidence nodes.

[0011] When generating a convergence signal, the path output unit is used to output the causal chain with the highest normalized probability.

[0012] Optional normalization processes include:

[0013] Keyframe extraction is performed on the acquired video stream;

[0014] Perform time-series interpolation on sensor logs;

[0015] Perform timeline mapping on the operation records.

[0016] Optionally, the process of constructing an event cause-effect graph using data alignment units includes:

[0017] Based on a pre-defined power system fault knowledge base, directed edges are established between initial evidence nodes.

[0018] Optionally, the calculation process for the chain likelihood score includes:

[0019] Obtain the node transition probability within the competitive causal chain;

[0020] Obtain the initial confidence level of the node;

[0021] Obtain the probabilities of competing causal chains in interpreting out-of-chain anomalous evidence;

[0022] The chain likelihood score is calculated by combining the node transition probability, initial confidence level, and explanatory probability.

[0023] Optionally, the fuzzy quantification unit employs a normalization function, and the normalized probability is calculated based on the chain likelihood score.

[0024] Optionally, the fuzzy quantification unit adopts the Shannon entropy formula and calculates the inter-chain fuzziness score based on the normalized probability.

[0025] Optionally, the evidence recognition unit may further be used to:

[0026] The evidence nodes in the event causal graph with an initial confidence level lower than a preset value or marked as missing are filtered to form a set of nodes to be verified.

[0027] The estimated information gain is calculated based on the set of nodes to be verified.

[0028] Optionally, the calculation process for estimating information gain includes:

[0029] Obtain the current inter-chain ambiguity score;

[0030] Simulation calculation of the new ambiguity when the nodes in the set of nodes to be verified are assumed to be true;

[0031] Simulation calculation of the new ambiguity when the node is assumed to be false;

[0032] Based on the current inter-chain ambiguity score, the new ambiguity when it is true, and the new ambiguity when it is false, the estimated information gain is calculated.

[0033] Optionally, the evidence identification unit is further used for:

[0034] Output key evidence points to the manual verification interface.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. This system unifies and normalizes heterogeneous asynchronous data such as videos and logs through data alignment and a fault knowledge base, constructing an event cause-effect graph that conforms to physical logic. This effectively solves the problem of difficulty in integrating multimodal data and provides a unified foundation for subsequent accurate reasoning.

[0037] 2. This system innovatively quantifies the ambiguity between multiple competing causal chains; by calculating the comprehensive likelihood score and the inter-chain ambiguity score, it transforms the uncertainty of tracing from a subjective concept into an objective and calculable indicator, thereby effectively distinguishing the credibility of different failure paths;

[0038] 3. This system addresses the pain point of the system being unable to guide verification when there is insufficient evidence or excessive ambiguity; when the system is confused, it can automatically calculate and identify the key evidence node that can reduce the uncertainty of the system, providing data-driven and highest-value guidance for manual verification;

[0039] 4. This system constructs an intelligent closed loop from data analysis, fuzzy quantification, key evidence identification to manual verification and feedback; it can automatically guide the convergence of the tracing process, avoid tracing interruptions caused by ambiguous evidence, and greatly improve the tracing efficiency and accuracy of complex power safety incidents. Attached Figure Description

[0040] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0041] Figure 1 is a structural diagram of the system of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0043] Example 1:

[0044] Please refer to Figure 1. The multimodal power safety protection equipment risk tracing system includes:

[0045] The data alignment unit is used to collect raw multimodal data and perform normalization processing to generate initial evidence nodes, and then construct an event causal graph.

[0046] The causal chain generation unit is used to search for competing causal chains based on the event causal graph and to calculate the chain likelihood score.

[0047] Fuzzy quantification unit, used to calculate normalized probability based on chain likelihood score and calculate inter-chain fuzziness score;

[0048] The convergence determination unit is used to compare the inter-chain ambiguity score with the preset narrative convergence threshold to obtain the convergence signal or the signal to be verified.

[0049] When a signal to be verified is generated, the evidence identification unit is used to calculate the estimated information gain and sort it to obtain the key evidence nodes.

[0050] When generating a convergence signal, the path output unit is used to output the causal chain with the highest normalized probability.

[0051] This invention provides a risk tracing system for power safety protection equipment based on multimodal data. The system aims to solve the problems of difficulty in tracing the cause and high ambiguity of the causal chain caused by multimodal heterogeneous data, asynchronous timing and loss of evidence after a power safety event such as a substation trip or line fault occurs.

[0052] As shown in Figure 1, the system specifically includes: a data alignment unit, a causal chain generation unit, a fuzzy quantification unit, a convergence determination unit, an evidence recognition unit, and a path output unit.

[0053] The data alignment unit aims to transform raw, heterogeneous, and asynchronous multimodal data into a structured probabilistic graphical model containing initial confidence levels, providing a unified data foundation for subsequent causal inference. In this embodiment, this unit collects raw multimodal data such as video streams, sensor logs, and operation records, and performs normalization processing on this data, such as timestamp alignment and semantic normalization. The processed data points are then converted into initial evidence nodes. The initial evidence node It refers to a structured data point that includes a timestamp. Data source Event Description and an initial confidence level The initial confidence level This refers to the initial assessment of the reliability of the evidence node, which is based on the reliability presets of the data source, such as physical sensor logs. The default value is 0.9, while inferred events based on video analysis The default value is 0.7; data is aligned to units and an event cause-effect graph is built based on these nodes;

[0054] The purpose of the causal chain generation unit is to search for all failure paths that may lead to the final risk event from a structured event causal graph and quantify the credibility of each path. In this embodiment, the unit locks an observed final risk event, such as the main transformer shutdown, as the target event in the graph. This unit is based on an event cause-effect graph from Start by performing a reverse graph search to find all possible competing causal chains. Where k=1...K; this unit calculates each chain Chain likelihood score ;

[0055] The purpose of the fuzzy quantification unit is to assess the overall uncertainty of the system regarding all K competing chains, that is, to quantify the degree of confusion of the system regarding which chain is true. In this embodiment, this unit is based on the likelihood scores of all chains calculated by the causal chain generation unit. Given a set of chains, calculate the normalized probability of each chain. To make their sum equal to 1; based on this probability distribution The unit calculates the inter-chain ambiguity score of the system. ;

[0056] The convergence determination unit aims to determine whether the current tracing results are clear and unambiguous enough, and whether further manual intervention is needed. In this embodiment, this unit is used to convert the inter-chain ambiguity score calculated by the ambiguity quantification unit. With the preset narrative convergence threshold Comparison; the narrative convergence threshold This refers to a scalar value, set through experimental calibration or expert experience, that can be tailored to the tolerance for uncertainty in a specific application scenario. For example, 0.5 bits represents the upper limit of acceptable uncertainty for the system. This indicates that the system is very certain that the probability of a certain chain is much higher than that of other chains, and at this time the unit generates a convergence signal; if This indicates system confusion, with multiple chains having comparable explanatory power, at which point the unit generates a signal to be verified;

[0057] The evidence identification unit aims to automatically identify the key data point that can most quickly and effectively reduce the level of confusion in the system when the system's ambiguity is too high, thus guiding manual verification. In this embodiment, this unit is activated only when it receives a signal from the convergence determination unit to generate a node to be verified. It is used to calculate the estimated information gain of certain nodes to be verified. After sorting, the key evidence nodes are obtained. ;

[0058] The path output unit aims to present the final and most credible fault cause chain to the user when the tracing task is successful, i.e., the ambiguity is sufficiently low. In this embodiment, this unit is activated only when it receives the convergence signal generated by the convergence determination unit; it is used to output the path with the highest normalized probability. That causal chain This serves as the final conclusion of this risk tracing process.

[0059] This embodiment constructs a closed-loop risk tracing system through the collaborative work of the above six units; it can not only identify possible causal chains like existing technologies, but also innovatively quantifies the ambiguity of evidence support between multiple competing chains. When the ambiguity is too high, automatically identify the key evidence node that can minimize the ambiguity. This solves the pain point of existing systems being unable to distinguish competing chains and guide the next step of verification when faced with multimodal heterogeneous data. It realizes an intelligent closed loop from data input to fuzzy quantification and then to key evidence feedback, which greatly improves the traceability efficiency and accuracy of complex power safety events.

[0060] Example 2:

[0061] Normalization includes:

[0062] Keyframe extraction is performed on the acquired video stream;

[0063] Perform time-series interpolation on sensor logs;

[0064] Perform timeline mapping on the operation records.

[0065] This embodiment is a specific implementation of the normalization process in the data alignment unit described in Embodiment 1; the purpose of this step is to unify data from different sources and in different formats onto the same spatiotemporal and semantic benchmark.

[0066] In this embodiment, the normalization process specifically includes:

[0067] Keyframe extraction from the acquired video stream: For example, the system processes the video stream using a preset target detection model such as YOLOv5 or a domain-specific power equipment model. When a change in the equipment status is detected, such as a switch opening or closing, an indicator light change, or an abnormality such as smoke or sparks, the system extracts the keyframe image at that moment and marks it as a specific event.

[0068] Perform time-series interpolation on sensor logs: For example, for sensor data with different sampling rates, such as temperature and voltage, use methods such as linear interpolation or spline interpolation to unify them to a standard time sampling rate, such as once per second, to ensure the comparability of different sensor data on the time axis.

[0069] Timeline mapping of operation records: For example, unstructured text in manually filled operation and maintenance logs, such as the execution of line A maintenance at 10:05, can be converted into structured events on a unified timeline through natural language processing and rule matching.

[0070] This embodiment achieves strict alignment and unification of heterogeneous data at both the temporal and semantic levels through the aforementioned specific processing methods; this ensures the accuracy of the initially generated evidence nodes. It has a consistent data format and comparable confidence levels. It is a necessary foundation for building high-quality, highly logical event cause-effect graphs, effectively avoiding subsequent reasoning errors caused by inconsistent data scales or disordered timing.

[0071] Example 3:

[0072] The process of constructing an event cause-effect graph using data alignment units includes:

[0073] Based on a pre-defined power system fault knowledge base, directed edges are established between initial evidence nodes.

[0074] This embodiment is a specific implementation of the process of constructing an event cause-effect graph in the data alignment unit described in Embodiment 1; the purpose of this process is to integrate discrete initial evidence nodes. Connect them into a graph structure that reflects cause and effect;

[0075] In this embodiment, the process is based on a preset power system fault knowledge base; the power system fault knowledge base refers to an expert rule set or graph database that contains the topological relationships of power system equipment, electrical wiring rules, and a large number of historical typical fault modes; its source is jointly constructed through domain expert knowledge and historical data mining;

[0076] At the initial evidence node After generation, the data alignment unit queries this knowledge base and establishes directed edges between the initial evidence nodes; the establishment of edges must simultaneously satisfy the following constraints:

[0077] Timing constraint: The timestamp of the cause node must be earlier than the timestamp of the result node;

[0078] Physical / Logical Constraints: The connection between two nodes must be supported by physical means such as electrical connection or logical means such as fault mode in the knowledge base; for example, even if nodes A and B are close in time, if the knowledge base shows that they are physically connected as belonging to two isolated subnets or logically connected as having no known fault association, then no edge is established.

[0079] This embodiment introduces a power system fault knowledge base as a strong constraint, ensuring that the constructed event causal graph conforms to the physical laws and business logic of the power industry. This greatly reduces pseudo-causal connections caused by similar time sequences, avoids combinatorial explosion in the graph model, and significantly improves the efficiency and search results of subsequent causal path search in Embodiment 1. The accuracy.

[0080] Example 4:

[0081] The calculation process of the chain likelihood score includes:

[0082] Obtain the node transition probability within the competitive causal chain;

[0083] Obtain the initial confidence level of the node;

[0084] Obtain the probabilities of competing causal chains in interpreting out-of-chain anomalous evidence;

[0085] The chain likelihood score is calculated by combining the node transition probability, initial confidence level, and explanatory probability.

[0086] This embodiment focuses on the chain likelihood score in the causal chain generation unit described in Embodiment 1. A concretization of the calculation process; the purpose of this score is to quantify each competing causal chain. The overall credibility of the evidence must take into account the logical rationality within the chain, the reliability of the evidence itself, and its explanatory power over external evidence.

[0087] In this embodiment, the chain likelihood score is calculated according to the following formula:

[0088] ;

[0089] in, Indicates the k-th chain The final likelihood score; Represents the traversal chain All evidence nodes ; This represents the node transition probability, i.e., given a parent node. Given the previous event on the chain, The probability of occurrence is obtained from the power system fault knowledge base described in Example 3, and is set based on historical fault data statistics or expert experience. The initial confidence level of the node is represented by the output of the data alignment unit in Example 1, which reflects the reliability of the evidence itself. Indicates all that do not belong to The set of off-chain abnormal evidence nodes; This represents the probability of interpreting out-of-chain anomalous evidence, i.e., under the assumption... Under the condition that the chain is true, observe off-chain evidence The probability is also obtained from the causal-phenomenon correlation matrix in the knowledge base;

[0090] This calculation process combines node transition probabilities, initial confidence levels, and explained probabilities to calculate the chain likelihood score;

[0091] The first part of the formula It's the core; it ensures the score. Weighted by two factors simultaneously:

[0092] The rationality of causal logic If the transition probability within a chain is very low, such as event A rarely leading to event B, It will decrease;

[0093] Reliability of the evidence itself If a chain of events, even if logically sound, depends on a low-confidence chain... Evidence nodes approaching zero are like a blurry video inference. It will also be significantly lowered;

[0094] This embodiment utilizes this comprehensive approach The calculation formula rigorously considers both the quality of evidence and causal logic; it ensures that the final likelihood score more accurately reflects the actual credibility of each causal chain, effectively addressing the potential for low-quality evidence chains in existing technologies. Low or weak logic chain The low-interference final judgment problem provides a more reliable input for subsequent fuzzy quantification.

[0095] Example 5:

[0096] The fuzzy quantification unit uses a normalization function to calculate the normalized probability based on the chain likelihood score; the fuzzy quantification unit uses the Shannon entropy formula to calculate the inter-chain fuzziness score based on the normalized probability.

[0097] This embodiment is a specific implementation of the fuzzy quantization unit described in Embodiment 1;

[0098] All were calculated in the embodiments. Likelihood score of each competitive chain After the set is assembled, the fuzzy quantization unit is activated, the purpose of which is to quantify this... The set may contain any positive value, which can be transformed into a scalar with a definite physical meaning that measures the overall uncertainty of the system. ;

[0099] Normalized probability calculation

[0100] The fuzzy quantification unit uses a normalization function, based on the chain likelihood score obtained in the previous step. Set, calculate the normalized probability of each chain. ;

[0101] Objective: The purpose of this step is to transform any positive value The set is transformed into a valid probability distribution with a sum of 1. This is to facilitate subsequent calculations of information entropy uncertainty, and the calculation formula is as follows:

[0102] ;

[0103] Let represent the normalized probability of the k-th chain, and its data type is a real number between 0 and 1. ; Indicates the first The likelihood score of the chain is obtained by calculation from the causal chain generating unit; Indicates all The sum of likelihood scores for each chain;

[0104] The effect of this step is to obtain a probability vector. It intuitively reflects the system's level of trust in each chain;

[0105] Inter-chain ambiguity calculation

[0106] The fuzzy quantization unit uses the Shannon entropy formula, based on the normalized probability obtained from the previous Softmax step. Distribution, calculation of inter-chain ambiguity scores in the system ;

[0107] Objective: The purpose of this step is to use Shannon entropy, a recognized mathematical metric in information theory, to objectively quantify the ambiguity of evidence supporting the system's understanding of the system's ambiguity. The degree of confusion in a vector distribution is calculated using the following formula:

[0108] ;

[0109] This represents the inter-chain ambiguity score, and its unit of measurement is bits. Indicates the first The normalized probability of the chain is obtained from the Softmax function calculated in the previous step of this unit;

[0110] This formula outputs a scalar. :

[0111] like For example, a line close to 0 A value close to 1 and the rest close to 0 indicates that the system is highly deterministic and the causal chain has converged.

[0112] like High, for example, multiple lines equal, achieve This indicates that the system is extremely confused, and multiple chains have comparable explanatory power;

[0113] This embodiment provides a solution based on the original likelihood score by combining Softmax and Shannon entropy. To the final system uncertainty The complete quantization path of bits with strict physical meaning makes ambiguity no longer a subjective concept, but an objective indicator that can be accurately calculated and compared, providing a reliable and quantitative decision basis for the subsequent convergence determination unit embodiment 1.

[0114] Example 6:

[0115] Before calculating the estimated information gain, the evidence recognition unit is further used for:

[0116] The evidence nodes in the event causal graph with an initial confidence level lower than a preset value or marked as missing are filtered to form a set of nodes to be verified.

[0117] The estimated information gain is calculated based on the set of nodes to be verified.

[0118] This embodiment is a detailed description of a pre-optimization operation performed by the evidence recognition unit described in Embodiment 1 before it performs its core function of calculating information gain;

[0119] After being activated by the verification signal from the convergence decision unit, the evidence identification unit's task is to find which node is most worthy of verification. Calculating all nodes in the graph would be extremely costly and unnecessary. For example, a... Nodes that have reached 0.99 do not require verification;

[0120] Therefore, in this embodiment, the unit is further used to perform an efficient screening step before calculating the estimated information gain:

[0121] This unit traverses all evidence nodes in the event cause-effect graph;

[0122] Filter out all initial confidence levels Lower than the preset value, for example, the business setting is... Or evidence nodes that were marked as missing during data collection;

[0123] These selected nodes form a set of nodes to be verified. ;

[0124] The set of nodes to be verified This refers to a subset containing all low-quality, uncertain, or missing evidence nodes, which are potential sources of confusion in the system.

[0125] The estimated information gain calculation will be based on and limited to this set of nodes to be verified. Instead of calculating for all nodes in the graph;

[0126] This embodiment significantly narrows the scope of information gain calculation by introducing this screening step before information gain calculation; it also significantly reduces the computational complexity of the evidence recognition unit and avoids errors in the calculation of already certain information gains. High-level nodes perform unnecessary assumption simulations, significantly improving the operational efficiency and response speed of the evidence recognition unit.

[0127] Example 7:

[0128] The calculation process for estimating information gain includes:

[0129] Obtain the current inter-chain ambiguity score;

[0130] Simulation calculation of the new ambiguity when the nodes in the set of nodes to be verified are assumed to be true;

[0131] Simulation calculation of the new ambiguity when the node is assumed to be false;

[0132] Based on the current inter-chain ambiguity score, the new ambiguity when it is true, and the new ambiguity when it is false, the estimated information gain is calculated.

[0133] This embodiment focuses on the estimated information gain in the evidence recognition unit described in Embodiment 1. The calculation process is specified; this calculation is performed after obtaining the set of nodes to be verified. Execute after;

[0134] The purpose of this calculation is to, for Each node in Simulation calculations show that if we send someone to verify... And obtained a definite result of true or false, and the total ambiguity of the system. How much will it decrease; this expected decrease is... Information gain In this embodiment, the calculation process for the estimated information gain follows the information gain formula:

[0135] ;

[0136] The calculation process includes:

[0137] Get the current inter-chain ambiguity score :

[0138] This represents the current inter-chain ambiguity score before verification, which is calculated by the ambiguity quantification unit in this iteration;

[0139] Simulation calculation of the new ambiguity when node j in the set of nodes to be verified is true. :

[0140] This represents the new ambiguity when the assumption that node j is true;

[0141] The method for obtaining this information is a simulation calculation: the system assumes that the confidence level of node j is... Forced to be set to 1.0; based on this new... The system then re-invokes the causal chain generation unit and the fuzzy quantification unit to calculate and obtain a new system fuzziness score, i.e. ;

[0142] Simulation calculation of the new ambiguity when node j is assumed to be false :

[0143] This represents the new ambiguity when the assumption that node j is false;

[0144] The acquisition method is similar; this is based on the system assumption that... After forcibly setting it to 0.0, the new system ambiguity score is recalculated, i.e. ;

[0145] Calculate the final gain:

[0146] This represents the prior probability of node j, which in this embodiment can be derived from its initial confidence level. ;

[0147] The system also uses the current inter-chain ambiguity score. The new fuzziness when it is true New fuzziness when it is false The estimated information gain of node j is calculated using the above formula. ;

[0148] This embodiment achieves [the desired result] through this assumption-simulation-recalculation approach. The precise quantification of the information value of each node to be verified; The physical unit is also bit. Consistent The higher the node, the stronger its ability to distinguish between K competing chains; this enables the system to automatically identify the node that, once verified, minimizes system confusion regardless of its authenticity, providing the most valuable, data-driven guidance for manual verification, and forming the core of the closed-loop feedback from fuzziness to convergence in the entire system.

[0149] Example 8:

[0150] The evidence identification unit is further used for:

[0151] Output key evidence points to the manual verification interface.

[0152] This embodiment is a refinement of a specific application output of the evidence recognition unit described in Embodiment 1 after recognizing key evidence;

[0153] Evidence recognition unit Calculate the information gain of all nodes in the set. And sorted in descending order, finding the one with the highest value node , i.e., key evidence nodes;

[0154] In this embodiment, the evidence identification unit is further configured to: identify the key evidence node The information is output to the manual verification interface;

[0155] The manual verification interface refers to a human-computer interaction interface used by power safety operation and maintenance personnel or risk analysts, such as a PC console or a mobile app.

[0156] The purpose of this output action is to transform the algorithm's analysis results into specific, executable manual verification instructions; for example:

[0157] when This is a low-confidence video inference event. For example, if the A disconnector switch is suspected of being activated at 10:02:05, the system will highlight it on the interface. The system cannot currently determine the cause. Please retrieve the original video of the A disconnector switch around 10:02:05 for manual confirmation first.

[0158] when If there is a missing sensor data point, the system can prompt you to check the historical data transmission of sensor B, or automatically trigger a higher-precision sensor data readback program.

[0159] This embodiment transforms the abstract algorithmic output of key evidence nodes into concrete, executable manual verification instructions and proactively pushes them to the operation and maintenance personnel's work interface, bridging the gap between intelligent data analysis and actual operation and maintenance. This ensures that the algorithm's intelligent analysis results can be utilized immediately and efficiently, avoiding a disconnect between analysis results and actual work, and greatly improving the overall efficiency of risk tracing. When the operation and maintenance personnel submit verification confirmation results through this interface, for example, by setting key evidence nodes... confidence level When updated to version 1.0 or 0.0, the system will use the updated set of evidence nodes to recalculate likelihood and fuzzy quantification starting from the causal chain generation unit, forming a complete convergent closed loop of human-computer interactive analysis-verification-reanalysis.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A risk tracing system for power safety protection equipment based on multimodal operation, characterized in that: include: The data alignment unit is used to collect raw multimodal data and perform normalization processing to generate initial evidence nodes, and then construct an event causal graph. The causal chain generation unit is used to search for competing causal chains based on the event causal graph and to calculate the chain likelihood score. Fuzzy quantification unit, used to calculate normalized probability based on chain likelihood score and calculate inter-chain fuzziness score; The convergence determination unit is used to compare the inter-chain ambiguity score with the preset narrative convergence threshold to obtain the convergence signal or the signal to be verified; when the signal to be verified is generated, the evidence identification unit is used to calculate the estimated information gain and sort it to obtain the key evidence nodes. When a convergence signal is generated, the path output unit is used to output the causal chain with the highest normalized probability. The chain likelihood score calculation process includes: obtaining the node transition probability within the competing causal chain; obtaining the initial confidence level of the node; obtaining the explanatory probability of the competing causal chain for out-of-chain anomalous evidence; and combining the node transition probability, initial confidence level, and explanatory probability to calculate the chain likelihood score; the explanatory probability is the probability of observing out-of-chain evidence under the assumption that the chain is true.

2. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 1, characterized in that, The normalization process includes: extracting keyframes from the acquired video stream; performing time-series interpolation on the sensor logs; and mapping the operation records to a timeline.

3. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 1, characterized in that, The process of the data alignment unit constructing the event causal graph includes: establishing directed edges between initial evidence nodes based on a preset power system fault knowledge base.

4. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 1, characterized in that, The fuzzy quantification unit uses a normalization function to calculate the normalized probability based on the chain likelihood score.

5. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 4, characterized in that, The fuzzy quantification unit uses the Shannon entropy formula to calculate the inter-chain fuzziness score based on normalized probability.

6. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 1, characterized in that, Before calculating the estimated information gain, the evidence identification unit is further used to: filter evidence nodes in the event causal graph whose initial confidence level is lower than a preset value or marked as missing, forming a set of nodes to be verified; the calculation of the estimated information gain is based on the set of nodes to be verified.

7. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 1, characterized in that, The process of calculating the estimated information gain includes: obtaining the current inter-chain ambiguity score; simulating the calculation of the new ambiguity when assuming the nodes in the set of nodes to be verified are true; simulating the calculation of the new ambiguity when assuming the nodes are false; and calculating the estimated information gain based on the current inter-chain ambiguity score, the new ambiguity when true and the new ambiguity when false.

8. The risk tracing system for power safety protection equipment based on multimodal operation according to claim 1, characterized in that, The evidence identification unit is further used to output key evidence nodes to the manual verification interface.

Images