Power production compliance judgment method and device and readable storage medium

Abstract

The invention discloses a power production compliance judgment method and device and a readable storage medium, and the method comprises the steps: carrying out the context association blocking of a system term binding structure on a system text, obtaining a system block text, carrying out the feature coding and priority classification of the system block text, and obtaining a system vector and a corresponding priority label; performing cross-modal linkage verification on the obtained multi-modal power production data to obtain a linkage verification result; carrying out vectorization and weighted fusion on the multi-modal power production data to obtain a fusion vector, carrying out similarity matching on the fusion vector and the system vector to obtain semantic similarity, and carrying out priority label weighting and threshold screening on the semantic similarity to obtain a matched system vector; and performing risk quantitative grading according to the linkage verification result and the priority label and the semantic similarity corresponding to the matching system vector to obtain an illegal risk grading result.

Description

Technical Field

[0001] This application relates to the field of power operation safety management, specifically to a method, device, and readable storage medium for judging power production compliance. Background Technology

[0002] The power production landscape is becoming increasingly complex and dynamic, with power grid operations becoming more flexible and operational procedures requiring greater precision. Therefore, strict adherence to various safety regulations is fundamental to ensuring power production safety in power production scenarios such as thermal power plants, new energy power plants, and energy storage dispatch centers. This necessitates real-time and accurate compliance assessments of power production operations.

[0003] Current compliance assessment methods use semantic segmentation algorithms to process power regulatory documents, dividing them into several regulatory clauses. Then, relevant operational data obtained during power production is used to match corresponding regulatory clauses through text similarity or cross-modal similarity for risk assessment, thereby determining whether there are any risks of violations in power production operations.

[0004] However, existing technologies, when segmenting regulatory clauses, rely solely on semantic coherence or single keywords, failing to consider specific binding structures within power regulations (such as the combination of strong constraints like "prohibiting live-line maintenance of 10kV and above equipment" with parameter thresholds). This leads to fragmented clauses, inaccurate extraction, and false alarms. Furthermore, existing technologies, relying solely on similarity matching, lack cross-system data linkage, making it difficult to effectively identify hidden violations across dimensions in power production operations (such as insufficient operator qualifications or prohibiting operation of equipment exceeding their qualification level), resulting in missed reports. Therefore, existing technologies, due to their difficulty in accurately extracting regulatory clauses and lack of cross-system data linkage, are unable to identify hidden violations, making them prone to false alarms and missed reports in power production compliance monitoring. Summary of the Invention

[0005] This application provides a method, device, and readable storage medium for judging compliance in power production, which can solve the technical problems in the prior art that make it difficult to accurately extract institutional clauses, lack cross-system data linkage, and fail to identify hidden violations, and are prone to false alarms and omissions in power production compliance monitoring.

[0006] In a first aspect, embodiments of this application provide a method for determining compliance in power production, including: The institutional text is segmented into institutional block texts based on the context association structure of the institutional clause binding structure. The institutional block texts are then subjected to feature encoding and priority classification to obtain institutional vectors and corresponding priority labels. Cross-modal linkage verification is performed on the acquired multimodal power production data to obtain the linkage verification results; Multimodal power production data is vectorized and weighted to obtain a fusion vector. The fusion vector is matched with the institutional vector to obtain semantic similarity. The semantic similarity is then weighted by priority labels and filtered by threshold to obtain the matching institutional vector. Based on the joint verification results and the priority labels and semantic similarity corresponding to the matching system vectors, risk quantification and classification are performed to obtain the violation risk classification results.

[0007] In conjunction with the first aspect, in one implementation method, the institutional text is divided into institutional block texts based on the contextual association structure of the institutional clause binding structure, including: The institutional text is segmented into words, and the segmentation type of each word is determined. When a combination of several consecutive word segmentation types meets the preset binding structure triggering conditions, adaptive association segmentation is performed based on the corresponding segmentation rules to obtain the system segmented text; The preset binding structure trigger conditions include level switching trigger conditions, threshold constraint trigger conditions, and scene operation trigger conditions.

[0008] In conjunction with the first aspect, in one implementation, cross-modal linkage verification is performed on the acquired multimodal power production data to obtain linkage verification results, including: Cross-modal device permission verification, device scenario verification, and operational load verification are performed on the acquired multimodal power production data to obtain the linkage verification results.

[0009] In conjunction with the first aspect, in one implementation, multimodal power production data is vectorized and weighted to obtain a fused vector, including: Multimodal power production data is vectorized based on preset vectorization rules for each modality to obtain multimodal vectors; The multimodal vector is obtained by weighting and fusing the multimodal vectors based on the grid load rate, power safety level, and preset modal weights.

[0010] In conjunction with the first aspect, in one implementation, feature encoding and priority classification are performed on the policy block text to obtain policy vectors and corresponding priority labels, including: Vectorization and semantic feature extraction are performed on the segmented text of regulations to obtain the regulation vector; Priority tags are obtained by keyword recognition and tag matching of the policy block text.

[0011] In conjunction with the first aspect, in one implementation, the priority label includes a constraint strength label, a policy clause level label, and a scenario classification label.

[0012] In conjunction with the first aspect, in one implementation, a matching vector is obtained by weighting semantic similarity using priority labels and applying a threshold, including: Weighted similarity is obtained by weighting semantic similarity based on constraint strength label, institutional clause level label, and scenario level label; The matching system vector is obtained by filtering the weighted similarity based on a preset similarity threshold.

[0013] In conjunction with the first aspect, in one implementation method, a risk quantification and classification result is obtained by performing risk classification based on the linkage verification result, the priority label corresponding to the matching system vector, and semantic similarity, including: The quantitative risk value is obtained by quantitative calculation based on the linkage verification results, the constraint strength label corresponding to the matching system vector, and the semantic similarity. The results of the violation risk classification are determined based on the quantified risk value and the preset classification threshold.

[0014] Secondly, embodiments of this application provide a power production compliance judgment device, comprising: The system clause block encoding module is used to divide the system text into blocks based on the context association of the system clause binding structure to obtain system block text, and to perform feature encoding and priority classification on the system block text to obtain system vectors and corresponding priority labels; The multimodal linkage verification module is used to perform cross-modal linkage verification on the acquired multimodal power production data to obtain the linkage verification results; The similarity matching module is used to vectorize and weightedly fuse multimodal power production data to obtain a fused vector, match the fused vector with the institutional vector to obtain semantic similarity, and perform priority label weighting and threshold filtering on the semantic similarity to obtain the matching institutional vector. The risk grading module is used to quantify and grade the risk based on the linkage verification results, the priority tags corresponding to the matching system vectors, and semantic similarity to obtain the violation risk grading results.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing a power production compliance judgment program, wherein when the power production compliance judgment program is executed by a processor, it implements the steps of the power production compliance judgment method as described in any of the above claims. The beneficial effects of the technical solutions provided in this application include: This application, through context-based segmentation based on the binding structure of institutional clauses, can effectively prevent the splitting of power institutional clauses during the segmentation process. By performing cross-modal linkage verification on multimodal power production data and cross-system data linkage, it can identify hidden violations in power production, thereby effectively avoiding false alarms or omissions in the power production compliance detection and judgment. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the power production compliance judgment method of this application; Figure 2 This is a flowchart illustrating the segmentation of the regulatory text in an embodiment of this application. Figure 3 This is a schematic diagram of the process for vectorizing multimodal power production data in this application; Figure 4 This is a schematic diagram of the system vector matching process in an embodiment of this application; Figure 5 This is a schematic diagram of the risk quantification and classification process in an embodiment of this application; Figure 6 This is a schematic diagram of the functional modules of an embodiment of the power production compliance judgment device of this application. Detailed Implementation

[0017] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0019] Firstly, embodiments of this application provide a method for determining compliance in power production.

[0020] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the power production compliance assessment method of this application. Figure 1 As shown, the methods for determining compliance in power production include: S101. The institutional text is divided into institutional block texts based on the context association structure of the institutional clause binding structure. The institutional block texts are then subjected to feature encoding and priority classification to obtain institutional vectors and corresponding priority labels. S102. Perform cross-modal linkage verification on the acquired multimodal power production data to obtain the linkage verification result; S103. Vectorize and adaptively weighted fuse the multimodal power production data to obtain a fusion vector. Match the fusion vector with the institutional vector to obtain semantic similarity. Perform priority label weighting and threshold filtering on the semantic similarity to obtain the matching institutional vector. S104. Based on the linkage verification results and the priority labels and semantic similarity corresponding to the matching system vector, risk quantification and classification are performed to obtain the violation risk classification result.

[0021] Specifically, in the embodiment of the power production compliance judgment method, it is first necessary to break down the various clauses in the policy text. Considering that conventional block division often uses fixed lengths or is based on "semantic coherence" or "single keyword" division, it is easy to segment certain specific binding structures in some power policies. In this regard, the embodiment adopts a context-related block division method based on the binding structure of policy clauses to divide the policy text into blocks, so as to ensure the integrity of the policy clause information in each block.

[0022] In the context-related segmentation based on the binding structure of institutional clauses, the embodiment first performs word segmentation on the institutional text and identifies the segmentation type of each word. By setting multi-level trigger conditions, when the segmentation types of several consecutive words meet any trigger condition, the text is then segmented according to the segmentation rules corresponding to that trigger condition. This ensures that specific binding structures in the institutional text are not split during the word segmentation process.

[0023] Meanwhile, for each policy clause obtained after segmentation, in order to ensure that the clauses with high importance are selected first during the matching process, the implementation example also needs to generate corresponding priority tags. In subsequent matching, the priority tags of each policy clause are used to ensure that important policy clauses are matched first.

[0024] Priority tags include constraint strength tags, rule / clause hierarchy tags, and scenario tags. Constraint strength tags are divided into strong and weak constraints; clauses related to personal, equipment, and power grid safety are strong constraints, while clauses related to process regulations are weak constraints. Rule / clause hierarchy tags indicate the level of the clause within the system, reflecting its "scope of application," and are divided into general provisions, specific provisions, and supplementary provisions. Scenario tags reflect the risk level of the working scenario, divided into high-risk scenarios and routine scenarios.

[0025] Then, the embodiment collects power production data through a multi-data source system and combines institutional clauses to determine whether there are any violations. At the same time, in order to identify latent violations in power production operations, which are difficult to be identified in single-modal data, it is necessary to analyze different-modal data. For this reason, the embodiment performs cross-modal linkage verification on the obtained multi-modal power production data to identify and analyze possible latent violations. The cross-modal linkage verification includes equipment permission verification, equipment scenario verification, and operation load verification. When any verification fails, it is considered that there is a violation.

[0026] In the process of matching institutional clauses, first, it is necessary to vectorize and fuse the production data of each modality. In the fusion process, considering that the sources of risks in different scenarios often vary, and correspondingly, the importance of the production data of each modality is also different, the embodiment sets a scenario-based weight to perform weighted fusion on the vectorized data of each modality to obtain a fusion vector. For example, in the coordinated scheduling of energy storage charging and discharging, it is necessary to consider time series and scheduling rules first. Therefore, text data and time series data have higher importance. Therefore, in the weighting process, these two types of data have higher weights.

[0027] Then, when matching and screening institutional clauses, the embodiment calculates the initial semantic similarity between the fusion vector and the institutional vector through cosine similarity. In the screening process, to ensure that important institutional clauses are matched first, the embodiment weights the semantic similarity through the constraint strength label, institutional clause hierarchy label, and scenario label of the institutional clause to obtain a weighted similarity, and obtains the matched institutional clauses according to the weighted similarity.

[0028] Finally, the embodiment performs a quantitative calculation on the linkage verification result, the constraint strength label corresponding to the matched institutional vector, and the semantic similarity to obtain a quantitative risk value, and classifies the quantitative risk value according to the set classification threshold to obtain a violation risk classification result.

[0029] In this embodiment, through the context-related chunking based on the institutional clause binding structure, it is possible to effectively avoid the splitting of power institutional clauses during the institutional clause chunking process. By performing cross-modal linkage verification on multi-modal power production data, cross-system data linkage is carried out to identify latent violations in power production, thereby effectively avoiding false positives or false negatives in the power production compliance detection and judgment. And when matching clauses, similarity weighting is performed through priority labels to ensure that important clauses are matched first. Finally, through quantitative calculation, the quantitative classification of violation risks is achieved.

[0030] Furthermore, in one embodiment Figure 2 is a schematic flowchart of the institutional text chunking of the embodiment of the present application, as Figure 2As shown, the institutional text is divided into institutional block texts by performing context-related segmentation based on the institutional clause binding structure, including: S201. Perform word segmentation on the institutional text and determine the word segmentation type of each word; S202. When a combination of several consecutive word segmentation types meets the preset binding structure triggering condition, adaptive association segmentation is performed based on the corresponding segmentation rules to obtain the system segmented text. The preset binding structure trigger conditions include level switching trigger conditions, threshold constraint trigger conditions, and scene operation trigger conditions.

[0031] Specifically, in the process of segmenting regulatory texts, traditional text segmentation methods, when processing power regulatory texts, do not consider some binding structures within the power regulations (such as the requirement to fully retain "strong constraint terms + voltage parameters" in "prohibition of live-line maintenance of 10kV and above equipment"). This can easily lead to the splitting of these binding structures during segmentation. Therefore, this embodiment uses context-based segmentation based on the binding structure of regulatory clauses to process the regulatory texts.

[0032] For the segmented processing of regulatory texts, the implementation example sets three levels of triggering conditions, including level switching triggering conditions, threshold constraint triggering conditions, and scenario operation triggering conditions. The level switching triggering condition is triggered when a change in chapter title (e.g., "General Provisions" → "Specific Provisions") and level keywords ("General Provisions", "Specific Provisions", "Supplementary Provisions") are detected; the threshold constraint triggering condition is triggered when a combination of a strong constraint word, a parameter word, and a threshold (e.g., "prohibited + voltage + ≥10kV", "must + maintenance cycle + ≤7 days") is detected; and the scenario operation triggering condition is triggered when a combination of a scenario word and an operation type (e.g., "peak summer demand + load adjustment", "Spring Festival power supply guarantee + closing operation") is detected.

[0033] The embodiment identifies the segmentation type of each segment after processing the institutional text into words. When a combination of segmentation types of several consecutive words meets the triggering condition (such as the third segment of a certain institutional text being a scene word and the fifth segment being an operation type), the corresponding segmentation rules are executed to obtain the institutional segmented text.

[0034] The block segmentation rules for triggering level switching are as follows: immediately interrupt the current block; the new block starts with the first 10 characters of the identifier; mark the level code; and merge blocks with fewer than 100 characters. This is to avoid semantic confusion between levels and ensure the coherence of clauses within a level.

[0035] For threshold constraint triggering conditions, the block segmentation rules are as follows: each block is independent; the complete parameter threshold is retained; a "strong constraint + parameter" label is added; and surrounding 50-character device terms are included. This avoids splitting core parameters and improves the matching accuracy of strong constraint clauses.

[0036] For scenario-based operation trigger conditions, the segmentation rules are as follows: independent segments; scenario-coded segments; and scenario-specific parameters (such as load rate thresholds) are associated. This achieves precise scenario-based matching and adapts to dynamic scenario requirements.

[0037] If the conditional block splitting rule is not triggered, then the following method is used for block splitting: initially split into blocks of 300 characters; check whether the block contains complete electricity clauses (including "Article X", "Section", "Item"); if incomplete, merge into adjacent blocks, and finally control the number of characters in the block to be between 150 and 500 characters to ensure semantic integrity.

[0038] In this embodiment, by using context-based segmentation based on the binding structure of institutional clauses, the electricity institutional clauses can be effectively prevented from being split during the segmentation process, thus ensuring the integrity of the braking clauses after segmentation.

[0039] Furthermore, in one embodiment, feature encoding and priority classification are performed on the policy block text to obtain policy vectors and corresponding priority labels, including: Vectorization and semantic feature extraction are performed on the segmented text of regulations to obtain the regulation vector; Priority tags are obtained by keyword recognition and tag matching of the policy block text.

[0040] Furthermore, in one embodiment, the priority label includes a constraint strength label, a policy clause level label, and a scenario classification label.

[0041] Specifically, for the segmented texts of regulations, the vectorization process is based on a BERT model pre-trained in the power sector. The model is based on a general BERT architecture and fine-tuned using three types of power-specific data: over 100,000 power regulation texts (including national regulations GB / T2900.50-2016, industry standard DL / T408-2018, and internal enterprise operation and maintenance procedures), over 50,000 cases of power violations (marked with the violations and related regulatory clauses), and over 30,000 power-related professional documents (covering equipment principles and operating specifications).

[0042] After fine-tuning, the model's accuracy in understanding power semantics is 35% higher than that of the general BERT, and it can accurately identify power-specific terms such as "GIS equipment," "SF6 gas leak detection," and "overload."

[0043] In addition, to give important clauses higher priority when matching institutional clauses, the implementation example also sets priority tags for each institutional block text, including constraint strength tags, institutional clause level tags, and scenario level tags.

[0044] In the tagging process, the implementation example identifies keywords in the policy block text, specifically strong constraint words, core terms related to power business, power parameters, power equipment, and power scenarios. For example, core terms for strong constraint words include "prohibited," "must," "strictly prohibited," and "must not," while core terms for power equipment include "GIS equipment," "main transformer," "switchgear," "energy storage battery pack," "photovoltaic inverter," and "interchange transformer." Then, the policy block text is categorized based on the identified keywords, and corresponding priority tags are assigned.

[0045] The constraint strength label includes strong constraints and weak constraints. Strong constraints include clauses related to personnel and equipment safety and core equipment operation, while weak constraints include clauses related to operating procedures and record-keeping standards.

[0046] The hierarchical labels for the regulations are divided into general provisions, specific provisions, and supplementary provisions, reflecting the scope of the provisions' effectiveness. The general provisions contain provisions that have global binding force, such as "This regulation applies to equipment of 10kV and above." The specific provisions include provisions related to specific operational requirements and parameter thresholds, such as "Maintenance interval before closing the circuit breaker must be greater than 7 days." The supplementary provisions include supplementary explanations, such as "Effective date of this regulation."

[0047] The scenario classification labels are divided into high-risk scenarios and routine scenarios, reflecting the risk level of the working conditions. The example classifies scenarios with high equipment load or prone to safety issues as high-risk scenarios, such as "peak summer power demand" and "Spring Festival power supply guarantee." Routine scenarios include general scenarios such as "routine inspections."

[0048] Furthermore, if the number of policy block texts is large, the implementation can also set the priority of the index based on the priority label when building the index. For example, for the constraint strength label, strong constraint clauses have a score of 1.5 when calculating the priority of the index, while weak constraint clauses have a score of 1.0, so as to ensure that strong constraint clauses are ranked higher in the index and are retrieved first.

[0049] In this embodiment, a BERT model pre-trained in the power sector is used to vectorize the policy block text. This vectorization process accurately identifies the semantic information of power-specific terms. By setting priority labels, policy conditions with higher importance are given higher priority when matching policy clauses, thereby improving the sensitivity to safety issues in the power production compliance judgment process.

[0050] Furthermore, in one embodiment, cross-modal linkage verification is performed on the acquired multimodal power production data to obtain the linkage verification result, including: Cross-modal device permission verification, device scenario verification, and operational load verification are performed on the acquired multimodal power production data to obtain the linkage verification results.

[0051] Specifically, in this embodiment, power production data from various data source systems are integrated and linked for verification to identify cross-modal latent violations.

[0052] The power production data includes equipment data, personnel data, and operational data. Equipment data is acquired from SCADA systems, equipment operation and maintenance platforms, and vibration monitoring systems, and includes data such as voltage, current, operating status, equipment maintenance cycles, and fault records. Personnel data is acquired through human resources systems, access control platforms, and safety training systems, and includes data such as job qualifications, operating authority scope, and power safety training records. Operational data is acquired through dispatch center systems, load monitoring platforms, and meteorological early warning systems, and includes data such as grid operation modes, historical loads, and meteorological early warnings.

[0053] Data acquired from different sources includes three types: time-series data, image data, and text data. To ensure data usability, preprocessing is required for each type. For time-series data (such as voltage, current, and vibration data), smoothing is performed to address potential noise and data fluctuations (e.g., electromagnetic interference during thunderstorms). For text data, standardized mapping of power terminology is necessary to address inconsistencies in textual representation across different systems.

[0054] To identify hidden violations in data from different modalities, the implementation example uses cross-modal device permission verification, device scenario verification, and operational load verification.

[0055] The equipment access control verification is defined as requiring the operator's qualification level to match the equipment voltage level, and ensuring the qualification is valid to prevent "low-qualified personnel from operating high-voltage equipment." Verification results are obtained by identifying the operator's qualification level in the personnel data and the equipment voltage level in the equipment data to perform qualification matching verification. A value of 1 indicates successful verification, while a value of 0 indicates failure. For example, if the operator's qualification level is "capable of operating 35kV equipment," but the equipment operates at 110kV, then the verification will fail.

[0056] Equipment scenario verification is defined as requiring equipment maintenance intervals to meet the condition of "base cycle × scenario weight," while ensuring the number of failures does not exceed a set limit, thus preventing "equipment exceeding maintenance limits from operating in high-risk scenarios." The base cycle is determined based on equipment manufacturing parameters, and the scenario weight is determined based on the safety risks of the equipment's operating scenario; for example, the scenario weight has increased adaptability for thunderstorm scenarios. The number of failures is used to prevent multiple minor failures from accumulating and causing safety accidents. The implementation example determines the scenario weight using business data and the equipment maintenance interval using equipment data. Verification is performed when the equipment maintenance interval is greater than the product of the base cycle and the scenario weight, and the number of failures is less than or equal to a set threshold. It is 1 if it is true, otherwise it is 0.

[0057] Operational load verification is defined as ensuring that the grid load rate does not exceed the limit and the voltage fluctuation amplitude is controllable after an operation, avoiding "grid overload or voltage collapse caused by the operation". This involves identifying operational behaviors in business data and calculating the load rate and voltage fluctuation amplitude after the operation based on equipment data. Verification is successful when both the load rate and voltage fluctuation amplitude after the operation are less than the set thresholds. The value is 1.

[0058] The final result is expressed as: In other words, the final verification is successful only if all three verifications are successful; otherwise, the verification fails and a corresponding warning is triggered.

[0059] In this embodiment, by performing cross-modal linkage verification on multimodal power production data and cross-system data linkage, hidden violations in power production are identified, effectively avoiding the underreporting of hidden violations in the power production compliance detection and judgment.

[0060] Furthermore, in one embodiment, Figure 3 This is a schematic diagram of the process for vectorizing multimodal power production data in this application, such as... Figure 3 As shown, the fusion vector is obtained by vectorizing and scene-adaptive weighted fusion of multimodal power production data, including: S301. Based on the preset vectorization rules of each modal data, the multimodal power production data is vectorized to obtain a multimodal vector; S302. The multimodal vectors are weighted and fused according to the power grid load rate, power safety level and preset modal weights to obtain the fused vector.

[0061] Specifically, in the process of vectorizing multimodal power production data, the implementation plan designs a power-specific vector generation method for three types of data: "text, time series, and image" to ensure that each vector can accurately capture power-specific information (rather than general features).

[0062] For text data, the implementation example also uses a pre-trained BERT model in the power sector for vectorization. During encoding, the implementation example incorporates three types of power-specific identifiers to ensure a strong binding between the vectors and the power scenario. For operation type identifiers, such as "closing operation," "equipment maintenance," and "energy storage charging / discharging," one-hot encoding is used to embed the vectors. Equipment number identifiers, such as "D-001 110kV main transformer" and "PCS-02 energy storage converter," are associated with equipment voltage level and model attributes. For time identifiers, accurate to the second, they are associated with the grid load period.

[0063] After encoding, a 768-dimensional text vector is generated, and the vector features can effectively preserve the identification information in the original text.

[0064] For time-series data, a time-specific TCN (Temporal Convolutional Network) is used for vectorization. The input layer consists of 30-dimensional time-series features, which are then processed through two convolutional layers, a pooling layer, and a fully connected layer to output a 768-dimensional vector. The first convolutional layer is used to capture short-term parameter mutations, the second convolutional layer enhances the identification of persistent anomalies, the pooling layer filters out transient noise, and finally, the fully connected layer transforms the time-series trend into a computable feature vector.

[0065] For image data, the implementation example uses a power safety equipment recognition model finely tuned to YOLOv8-tiny. The model focuses only on five essential power safety equipment categories (insulating gloves, safety helmets, grounding wires, insulating boots, and voltage detectors), excluding irrelevant items (such as ordinary gloves and work clothes). It employs binary tag encoding, with each category of equipment corresponding to one dimension: "compliance = 1, non-compliance = 0". The output is a 5-dimensional "safety equipment compliance vector". For example, if an operator in an image is wearing a safety helmet but not insulating boots, and a grounding wire is connected, the recognition result would be vector [1,0,1,0,0] (dimension order: safety helmet, insulating gloves, grounding wire, insulating boots, voltage detector).

[0066] For the three types of vectors obtained above, the embodiment designs a weighted fusion method based on safety priority, rather than the general equal-weight fusion, to ensure that the vectors are tilted towards the core elements of power safety. The weighted fusion formula is expressed as:

[0067] in, This represents a text vector with a weight of 0.5. Represents a time-series vector. Given the current grid load factor, Represents an image vector. The power safety level is determined based on the safety risks of the current operation and is issued in real time by the dispatch center. The safety level includes three levels: high, medium, and low, with corresponding values ​​of 1.5, 1.0, and 0.8. The higher the safety level, the greater the weight of the image vector, in order to strengthen the compliance verification of power operation equipment.

[0068] In this embodiment, different vectorization rules are set for data of different modalities to ensure that each vector can accurately capture power-specific information rather than general features. Furthermore, a weighted fusion method based on security priority is adopted during fusion to ensure that the data is biased towards core elements of power security.

[0069] Furthermore, in one embodiment, Figure 4 This is a schematic diagram of the system vector matching process in an embodiment of this application, as shown below. Figure 4 As shown, the matching system vector is obtained by weighting semantic similarity by priority labels and thresholding, including: S401. Weighted similarity is obtained by weighting semantic similarity based on constraint strength label, institutional clause level label and scenario level label. S402. Based on a preset similarity threshold, the weighted similarity is filtered to obtain the matching system vector.

[0070] Specifically, in the process of matching institutional segmented text, the example first calculates the cosine similarity between the fused vector and each institutional vector:

[0071] in, The cosine similarity between the fusion vector and the institutional vector reflects the basis of semantic matching. Represents the fusion vector. Indicates the first A system vector.

[0072] Then, in order to prioritize the matching of important policy clauses during the screening process, the example uses priority tags of policy vectors to weight semantic similarity:

[0073] in, Indicates weighted similarity. This indicates the binding strength of the institutional clauses (1.5 for strong constraints and 1.0 for weak constraints). The hierarchical labels indicate the level of the institutional provisions (1.2 for general provisions, 1.0 for specific provisions, and 0.8 for supplementary provisions). The scenario classification labels indicate the braking clauses (1.2 for high-risk scenarios and 1.0 for normal scenarios).

[0074] The implementation then filters the weighted similarity using a similarity threshold T, deleting invalid matches with a weighted similarity less than the threshold T. The value of the similarity threshold T can be determined based on the current safety level of the power operation. The safety level is determined by setting different thresholds, for example, 0.8, 0.65, and 0.5 from highest to lowest. This embodiment sets different screening thresholds for different safety levels to accommodate the safety requirements of different work conditions. The safety level is issued in real time by the dispatch center based on the current work type.

[0075] After filtering, the remaining policy clauses are sorted according to weighted similarity, and the top 5 clauses with the highest similarity are output. If no clause meets the threshold, a no-match policy warning or a downgraded search strategy is triggered.

[0076] In this embodiment, when matching clauses, similarity weighting is performed by priority tags, which can ensure that important clauses are matched first and improve the detection rate of high-risk violations.

[0077] Furthermore, in one embodiment, Figure 5 This is a schematic diagram of the risk quantification and classification process in an embodiment of this application, as shown below. Figure 5 As shown, based on the linkage verification results and the priority labels and semantic similarity corresponding to the matching system vectors, risk quantification and classification are performed to obtain the violation risk classification results, including: S501. Based on the linkage verification results and the constraint strength label and semantic similarity corresponding to the matching system vector, the quantitative risk value is obtained by quantitative calculation. S502. Determine the violation risk classification result based on the quantified risk value and the preset classification threshold.

[0078] Specifically, after obtaining the matching system vector, the implementation example calculates the risk value based on the following formula:

[0079] in, This indicates the result of the linkage verification. The value is 1.0 when the verification passes and 0.5 when the verification fails. The score for this item is halved, which directly leads to a decrease in the overall compliance and thus increases the risk value.

[0080] This indicates the binding strength of the policy terms, reflecting the level of risk when the policy terms are violated, and is ensured to be in the range [0,1] by dividing by the maximum value of 1.5.

[0081] To achieve the basic cosine similarity between the fusion vector and the institutional vector, for the braking clauses obtained through screening and matching, the similarity will decrease when there is a violation of the institutional clauses, so it can be used as an indicator for calculating the risk value.

[0082] To normalize the safety level, the safety center issues and fine-tunes the calculated risk values ​​to adjust the scores for different work scenarios. The values ​​are then divided by the maximum value of 1.5 to ensure they fall within the range of [0,1].

[0083] Based on the calculated quantitative risk values, the implementation example sets multiple risk levels. A risk value greater than or equal to 0.3 is classified as high risk, potentially indicating a violation of strong constraints and ineffective linkage, which could easily lead to safety accidents. Typical scenarios include "unqualified operation of high-voltage equipment, substandard maintenance, or heavy-load operation." A risk value greater than or equal to 0.1 and less than 0.3 is classified as medium risk, potentially indicating a violation of weak constraints, potential risks, or equipment malfunctions. Typical scenarios include "over-limit load adjustments" and "missing safety equipment." A risk value less than 0.1 is classified as low risk, requiring attention to any non-compliant behavior. A typical scenario is "compliant procedures but improper equipment storage."

[0084] Furthermore, the implementation plan includes different early warning strategies and rectification timelines for different risk scenarios. For example, for high-risk situations, early warnings are issued via "dispatch pop-up + voice call + security director terminal + operations and maintenance APP," with a rectification timeline of less than 2 hours; for medium-risk situations, early warnings are issued via "operations and maintenance APP + team leader terminal + monitoring platform," with a rectification timeline of 24 hours; and for low-risk situations, early warnings are issued via "operations and maintenance APP + monitoring platform," with a rectification timeline of 48 hours. In addition, different verification methods are specified for different risk levels. Early warning information includes the early warning number, basic operational information, violation basis (clause + linkage result), risk value, rectification requirements, and supporting data links, ensuring complete and unambiguous information.

[0085] Secondly, embodiments of this application also provide a power production compliance judgment device.

[0086] In one embodiment, reference is made to Figure 6 , Figure 6 This is a functional module diagram of an embodiment of the power production compliance judgment device of this application. Figure 6 As shown, the power production compliance assessment device includes: The system clause block encoding module 601 is used to divide the system text into blocks based on the context association of the system clause binding structure to obtain system block text, and to perform feature encoding and priority classification on the system block text to obtain system vector and corresponding priority label; The multimodal linkage verification module 602 is used to perform cross-modal linkage verification on the acquired multimodal power production data to obtain the linkage verification result; The similarity matching module 603 is used to vectorize and weightedly fuse multimodal power production data to obtain a fused vector, match the fused vector with the institutional vector to obtain semantic similarity, and perform priority label weighting and threshold filtering on the semantic similarity to obtain a matching institutional vector. The risk grading module 604 is used to quantify and grade the risk based on the linkage verification results, the priority tags corresponding to the matching system vectors, and semantic similarity to obtain the violation risk grading results.

[0087] The functions of each module in the aforementioned power production compliance judgment device correspond to the steps in the aforementioned power production compliance judgment method embodiment, and their functions and implementation processes will not be described in detail here.

[0088] Thirdly, embodiments of this application also provide a computer-readable storage medium.

[0089] The present application has a computer-readable storage medium storing a power production compliance judgment program, wherein when the power production compliance judgment program is executed by a processor, it implements the steps of the power production compliance judgment method described above.

[0090] The method implemented when the power production compliance judgment procedure is executed can be referred to in the various embodiments of the power production compliance judgment method of this application, and will not be repeated here.

[0091] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0092] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0093] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0094] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0095] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods of the various embodiments of this application.

[0097] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for judging compliance in power production, characterized in that, include: The policy text is segmented into policy block texts based on the context association structure of the policy clause binding structure. The policy block texts are then subjected to feature encoding and priority classification to obtain policy vectors and corresponding priority labels. Cross-modal linkage verification is performed on the acquired multimodal power production data to obtain the linkage verification results; The multimodal power production data is vectorized and weighted to obtain a fusion vector. The fusion vector is matched with the institutional vector to obtain semantic similarity. The semantic similarity is then weighted by priority labels and filtered by threshold to obtain a matching institutional vector. Based on the linkage verification results and the priority labels and semantic similarity corresponding to the matching system vector, risk quantification and classification are performed to obtain the violation risk classification result.

2. The method for judging compliance in power production according to claim 1, characterized in that, The process of dividing the policy text into policy block texts based on the context association structure of policy clause binding to obtain policy block text includes: The institutional text is segmented into words, and the segmentation type of each word is determined. When a combination of several consecutive word segmentation types satisfies the preset binding structure triggering condition, adaptive association segmentation is performed based on the corresponding segmentation rules to obtain the system segmented text; The preset binding structure triggering conditions include level switching triggering conditions, threshold constraint triggering conditions, and scene operation triggering conditions.

3. The method for judging compliance in power production according to claim 1, characterized in that, The process of performing cross-modal linkage verification on the acquired multimodal power production data to obtain linkage verification results includes: Cross-modal device permission verification, device scenario verification, and operational load verification are performed on the acquired multimodal power production data to obtain the linkage verification results.

4. The method for judging compliance in power production according to claim 1, characterized in that, The process of vectorizing and weighting the multimodal power production data to obtain a fused vector includes: The multimodal power production data is vectorized based on the preset vectorization rules of each modal data to obtain multimodal vectors; The multimodal vectors are weighted and fused according to the grid load rate, power safety level, and preset modal weights to obtain a fused vector.

5. The method for judging compliance in power production according to claim 1, characterized in that, The step of performing feature encoding and priority classification on the segmented text of the regulations to obtain the regulation vector and corresponding priority label includes: The institutional vector is obtained by vectorizing the segmented text of the system and extracting its semantic features. Priority tags are obtained by keyword recognition and tag matching of the segmented text of the system.

6. The method for judging compliance in power production according to claim 1, characterized in that, The priority tags include constraint strength tags, institutional clause level tags, and scenario classification tags.

7. The method for judging compliance in power production according to claim 6, characterized in that, The process of obtaining a matching system vector by prioritizing and weighting semantic similarity with a threshold includes: A weighted similarity is obtained by weighting the semantic similarity based on the constraint strength label, the institutional clause level label, and the scenario classification label. The matching system vector is obtained by filtering the weighted similarity based on a preset similarity threshold.

8. The method for determining compliance in power production according to claim 6, characterized in that, The step of obtaining a violation risk classification result by performing risk quantification and classification based on the linkage verification result and the priority label and semantic similarity corresponding to the matching system vector includes: The quantitative risk value is obtained by quantitative calculation based on the linkage verification result, the constraint strength label and semantic similarity corresponding to the matching system vector; The violation risk classification result is determined based on the quantified risk value and the preset classification threshold.

9. A power production compliance judgment device, characterized in that, include: The institutional clause block encoding module is used to divide the institutional text into institutional block texts based on the context association structure of the institutional clause binding structure, and to perform feature encoding and priority classification on the institutional block texts to obtain institutional vectors and corresponding priority labels. The multimodal linkage verification module is used to perform cross-modal linkage verification on the acquired multimodal power production data to obtain the linkage verification results; The similarity matching module is used to vectorize and weightedly fuse the multimodal power production data to obtain a fused vector, match the fused vector with the institutional vector to obtain semantic similarity, and perform priority label weighting and threshold filtering on the semantic similarity to obtain a matching institutional vector. The risk grading module is used to perform risk quantification and grading based on the linkage verification results and the priority tags and semantic similarity corresponding to the matching system vector to obtain the violation risk grading results.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a power production compliance judgment program, wherein when the power production compliance judgment program is executed by a processor, it implements the steps of the power production compliance judgment method as described in any one of claims 1 to 8.

Images