Work ticket safety control identification method and system based on artificial intelligence
By automatically identifying and analyzing work orders using artificial intelligence technology, and combining multimodal attention and self-learning mechanisms, the problem of low efficiency and easy omissions in traditional manual review has been solved, achieving fast, accurate and adaptive security management of work orders.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-03
AI Technical Summary
Traditional work order security management relies on manual review, which is inefficient, prone to omissions, lacks dynamic risk prediction and multi-source data fusion capabilities, and cannot adapt to dynamic changes in work order content, resulting in poor system generalization ability.
An artificial intelligence-based approach is adopted, which combines image processing and multimodal attention mechanisms with the MaxKB knowledge base management system and GLM-4V-Flash visual encoder to automatically identify text and visual features in work orders, establish a comprehensive rule base for multi-round reasoning, and achieve self-learning and dynamic updates.
The system automates and enables rapid review of work orders, improving review efficiency and accuracy, ensuring comprehensiveness and adaptability of security checks, reducing missed checks and misjudgments, and continuously optimizing the system to adapt to new scenarios.
Smart Images
Figure CN121600538A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system management technology, and in particular to a work ticket security control and identification method and system based on artificial intelligence. Background Technology
[0002] In the operation of power systems, work permits are written orders authorizing work on electrical equipment and system software, and are also the core basis for implementing safety technical measures. Traditional work permit safety management mainly relies on manual review, with safety supervisors checking each item on the work permit for proper completion, completeness of safety measures, and logical time consistency. However, with the expansion of power marketing field operations and the increasing prevalence of small, scattered, and urgent tasks, the number of operations has surged, revealing the following serious shortcomings of the manual review method: First, manual review is inefficient, taking an average of about 30 minutes to review each work order, which cannot meet the needs of real-time safety management. In particular, delays in emergency repair operations may lead to safety hazards.
[0003] Second, manual review is prone to omissions, especially high-risk items such as inconsistencies in time logic, mismatches between safety measures and work content, and lack of standardized signatures.
[0004] Third, most existing electronic work order systems only digitize forms and lack the ability to dynamically predict risks and integrate multi-source data. For example, rule-based systems only support fixed template validation and cannot adapt to dynamic changes in work order content; pure optical character recognition solutions ignore the semantic relationship between text and images, resulting in a high false positive rate.
[0005] Fourth, the existing system lacks self-learning and knowledge evolution mechanisms, and cannot adaptively update according to new operational risks or rule changes, resulting in poor system generalization ability and decreased accuracy after long-term use.
[0006] Therefore, there is an urgent need for a technical solution that can automatically, in real time, and accurately identify security risks in work tickets and has continuous learning capabilities to solve the above problems. Summary of the Invention
[0007] To achieve the above objectives, the present invention provides a work order security management and identification method based on artificial intelligence, the method comprising the following steps: Step 1: Collect written work orders for on-site operations of the power system, convert the PDF work orders into an image format that the system can recognize, and perform resolution normalization, noise removal and image enhancement processing on the images; Step 2: Construct a work permit security inspection rule base, which includes rules for checking misuse of permit types, non-standard work tasks, time-related issues, three types of personnel and signature issues, safety measures issues, and dual-permission issues. Step 3: Parse the image generated in Step 1 using the MaxKB knowledge base management system to extract the text feature vector from the work order, and simultaneously extract the visual feature vector from the work order image using the GLM-4V-Flash visual encoder; Step 4: Input the text feature vector and visual feature vector into the cross-modal attention layer. Map the text feature vector and visual feature vector to the same semantic space through a learnable projection matrix. Calculate the association weight between the visual feature vector and the text feature vector using a multi-head attention mechanism to generate a fused feature vector. Step 5: Based on the fused feature vector, the GLM-4V-Flash logic reasoning engine is combined with the rule base of Step 2 to perform multi-round reasoning and output the normative judgment result of the work order. The judgment result includes the specific problem type of meeting or not meeting the requirements and the correction suggestions. Step 6: Based on the judgment results of Step 5, dynamically update the MaxKB knowledge base and the fine-tuning dataset of the GLM-4V-Flash visual encoder. New rules are written into the rule base after manual review, and visual recognition error cases are added to the training dataset after being labeled.
[0008] Preferably, the rules for checking ticket misuse in step 2 specifically include: Based on the work tasks described on the work order, it is determined that the work tasks involve a 35 kV voltage level corresponding to a power supply work order, the work tasks involve a 20 kV voltage level corresponding to a power distribution work order, and the work tasks involve a 10 kV and below voltage level corresponding to a low-voltage work order. If the type of work ticket does not match the content of the work task, it will be marked as a problem of incorrect ticket type.
[0009] Preferably, the rules for checking non-standard work tasks in step 2 specifically include: Analyze the text content of the work task field in the work order. If the text content contains the suffix "installation, disassembly and replacement" and does not belong to the "metering device failure_installation, disassembly and replacement" type of task, it is determined that the work task description is unclear. The system maintains a list of common error descriptions, including "low-voltage new installation / capacity increase," "metering equipment replacement," and "account cancellation." The system performs real-time matching between text feature vectors and the error description list.
[0010] Preferably, the time issue checking rules in step 2 specifically include: Verify the logical relationships between the planned start time, planned end time, issuance time, permission time, termination time, and extension time in the work order, and verify that the issuance time is earlier than the planned start time, the permission time is later than the planned start time, the termination time is earlier than the planned end time, the extension time is earlier than the planned end time, and the validity period of the work order does not span days. The time arrangement is judged based on the number of team members and the workload. If the task time for a single person to replace 150 meters is less than the preset threshold, it is marked as a time problem.
[0011] Preferably, the specific rules for checking the three types of people and signatures in step 2 include: By analyzing the pixel distribution of the signature area using visual feature vectors, handwritten signatures can be distinguished from machine-printed signatures. Except for telephone authorization scenarios, the licensor's signature must be handwritten; if a machine-printed signature is detected, it is considered a signature problem. No judgment is needed if the "Work Supervisor" and "Work Team Member" fields do not contain a signature area.
[0012] Preferably, the safety measure issue inspection rules in step 2 specifically include: Dynamically match safety measures to the work content, and check whether the safety measures are consistent with the work content, whether they fully cover the risk points, and whether they are targeted. If the field for the live parts to be retained in the live work task is "None" or the field for the live parts to be retained in the meter box replacement task is "All Retained", then it is marked as a safety issue. Verify whether the safety measures include voltage testing procedures, requirements for wearing safety goggles, measures to prevent short circuits or grounding in the voltage secondary circuit, and measures to prevent open circuits in the current secondary circuit.
[0013] Preferably, the dual-licensing issue checking rules in step 2 specifically include: By combining the information on the work order with the data from the on-site work platform, it can be determined whether dual authorization is required for the work on the customer side. If the work order is a low-voltage work order or a transformer task, then dual authorization is not required; If a work order involves a non-low-voltage task and dual licensing is not implemented, it is marked as a dual licensing issue.
[0014] Preferably, the text feature extraction and visual feature extraction in step 3 specifically include: The MaxKB knowledge base management system performs word segmentation, entity recognition, and relation extraction on work order texts, and outputs text feature vectors. The GLM-4V-Flash visual encoder extracts image features based on the ViT architecture, including signature areas, table structures, and handwritten text areas, and outputs visual feature vectors. Textual and visual feature vectors are fused through a multi-level extraction pipeline to generate structured data.
[0015] Preferably, the cross-modal attention layer in step 4 specifically includes: Text feature vectors and visual feature vectors are mapped to a 512-dimensional space using linear projection matrices; The mapped text feature vector and visual feature vector are input into a multi-head attention mechanism to calculate the association weights between the query vector, key vector, and value vector. The association weights are used to weight and sum the text feature vectors and visual feature vectors to generate a fused feature vector. Step 6, the dynamic updating of the knowledge base, specifically includes: New rules are written into the MaxKB knowledge base after being manually reviewed; Visual recognition error cases were annotated and added to the fine-tuning dataset of the GLM-4V-Flash visual encoder; The system periodically retrains the projection matrix in the cross-modal attention layer to adapt to changes in feature distribution.
[0016] An AI-based work ticket security management and identification system includes a memory configured to store instructions, a processor configured to retrieve the instructions from the memory, and an AI-based work ticket security management and identification method implemented when executing the instructions.
[0017] The beneficial effects of this invention are: 1. This invention combines image processing and artificial intelligence technologies to automate the analysis and review of work orders. The system can complete the standardization check of work orders in a short time, avoiding the inefficiency of manual review. Especially when tasks are urgent, it can quickly identify problems and reduce the time spent on manual review. Through automated processing, the efficiency of work order review is greatly improved, enabling real-time response to safety control requirements.
[0018] 2. This invention establishes a comprehensive rule base covering multiple inspection items such as ticket misuse, task compliance, and time logic. It extracts comprehensive text and visual feature vectors through the MaxKB knowledge base management system and the GLM-4V-Flash visual encoder, and performs precise analysis using a multi-round inference engine, ensuring that no high-risk items are missed. The system can automatically detect and mark non-compliant work ticket content, significantly improving the coverage of missed detections and ensuring the comprehensiveness and accuracy of security checks.
[0019] 3. This invention combines text and image information, employing a cross-modal attention mechanism to deeply fuse text features with visual feature vectors, significantly improving the semantic correlation between text and images. By introducing a self-learning mechanism, the system can dynamically adapt to changes in work order content, intelligently analyze different types of work orders, reduce misjudgments and omissions, and improve the system's accuracy and adaptability.
[0020] 4. This invention achieves adaptive learning of the system by dynamically updating the MaxKB knowledge base and combining manual review and visual error labeling mechanisms. New rules and error cases are added to the training dataset through manual review and labeling, driving continuous system optimization and enhancing its generalization ability. By periodically retraining the projection matrix in the cross-modal attention layer, the system can adapt to new feature distributions, thereby ensuring long-term stable operation of the system and avoiding accuracy degradation. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a flowchart illustrating the steps of the safety measure problem checking rules in step 2 of the method of the present invention; Figure 3 This is a flowchart illustrating the steps of checking the dual-licensing issue rules in step 2 of the method of the present invention. Detailed Implementation
[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.
[0024] Please see Figures 1-3 This invention provides an artificial intelligence-based method for safety management and identification of work permits. Specifically, it first collects written work permits for on-site operations in the power system, which are typically stored in PDF format. The system converts the PDF work permits into a recognizable image format and performs resolution standardization processing on the images to ensure that all work permit images have consistent pixel density.
[0025] Furthermore, image denoising algorithms are used to eliminate interference noise generated during scanning or shooting, and image enhancement techniques are employed to optimize image contrast and brightness, making text, tables, and signature areas clearer. This process provides a high-quality data foundation for subsequent text recognition and visual feature extraction.
[0026] Furthermore, a complete work order security check rule base was constructed, covering rules for incorrect use of work orders, non-standard tasks, time scheduling, signature issues, security measures, and dual authorization. The MaxKB knowledge base management system was used to parse images and extract text feature vectors, while a GLM-4V-Flash visual encoder was used to extract visual feature vectors. A cross-modal attention layer was used to map the text and visual feature vectors to the same semantic space, and a multi-head attention mechanism was used to calculate the association weights between the two, generating a fused feature vector.
[0027] Finally, the system integrates feature vector input logic reasoning engines, performs multi-round reasoning with a rule base, and outputs standardized judgment results for work orders, providing specific problem types and correction suggestions. The system dynamically updates the knowledge base and fine-tunes the dataset based on the reasoning results. Newly added rules are written into the knowledge base after manual review, and visual recognition error cases are annotated and added to the training dataset for self-learning optimization.
[0028] Automated processing of work orders significantly improves review efficiency and accuracy, reducing manual operation time and human error. Cross-modal analysis methods integrating text and visual features enable a more comprehensive understanding of work order content, ensuring effective identification of various safety issues. A dynamic update mechanism ensures the system continuously learns and adapts to new tasks and scenarios, improving the safety and standardization of power operations.
[0029] In one possible implementation, the system first parses the task description field of the work order to identify the voltage level involved in the task. If the task involves a 35 kV voltage level, a power distribution work order is used; if it involves a 20 kV voltage level, a distribution work order is used; and if it involves a 10 kV or lower voltage level, a low-voltage work order is used. The system automatically compares the extracted task content with the work order type. If a mismatch is found, it is marked as a misuse of work order type. Simultaneously, the system records this misuse case in the knowledge base for subsequent rule optimization. This rule can perform batch checks on a large number of work orders during real-time operation, improving the efficiency of work order type approval.
[0030] This invention ensures a high degree of matching between work order types and actual work tasks, effectively avoiding safety risks caused by incorrect use of work orders. The automated matching mechanism reduces subjective errors in human judgment, ensuring standardized work order procedures during power operations and improving operational safety.
[0031] In one possible implementation, the system first analyzes the text content of the task field on the work order in its rules for checking non-standard work tasks. If the text content contains the suffix "installation, dismantling, and replacement" but does not belong to the "metering device failure_installation, dismantling, and replacement" category, it is determined that the work task description is unclear. The system maintains a list of common erroneous descriptions, including "low-voltage new installation and capacity expansion_installation, dismantling, and replacement", "metering equipment replacement_installation, dismantling, and replacement", and performs real-time matching with this list using text feature vectors. The matching results are used to determine whether the task description is standardized. If non-standard descriptions are found, they are marked as task description problems, and correction suggestions are provided in the inference results. The erroneous descriptions can be updated periodically to adapt to changes in different scenarios and job types.
[0032] The embodiments of this invention ensure that work tasks are clearly and accurately described, reducing operational errors and safety hazards caused by ambiguous descriptions. Real-time matching and error list maintenance mechanisms improve the system's adaptability and accuracy, enabling more refined task review.
[0033] In one possible implementation, the time-related issue checking rules verify the logical relationships between the planned start time, planned end time, issuance time, permission time, termination time, and extension time in the work order. Specific rules include an issuance time earlier than the planned start time, a permission time later than the planned start time, a termination time earlier than the planned end time, and an extension time earlier than the planned end time, ensuring that the work order validity does not span multiple days. The system also considers the number of work team members and the workload to determine the reasonableness of the time arrangement. For example, if the task time for a single person to replace 150 meters is less than a preset threshold, it is marked as a time issue. This rule can be compared with time thresholds and work standard data in the knowledge base, automatically outputting a reasonableness judgment.
[0034] By automatically verifying the logical relationships between time items and the allocation of tasks, the system ensures that work order time arrangements are reasonable, avoiding accidents caused by overloading or underestimating the time required for tasks. This rule improves the standardization of time management and guarantees the safe and efficient completion of work.
[0035] In one possible implementation, among the three signature issue checking rules, the system analyzes the pixel distribution of the signature area using visual feature vectors to distinguish between handwritten and machine-printed signatures. Except for telephone authorization scenarios, the authorizing person's signature must be handwritten; if a machine-printed signature is detected, it is considered a signature issue. For the "Work Supervisor" and "Work Team Member" fields, if they do not contain a signature area, no judgment is needed. The system compares the analysis results with signature specification requirements in the rule base, outputs signature issue markers and correction suggestions, and error cases can be fed back to the knowledge base to optimize signature recognition capabilities.
[0036] This rule ensures that signatures are authentic and valid, preventing security and liability risks caused by forged or improper signatures. Automated signature detection improves review efficiency and guarantees the compliance of work permits.
[0037] In one possible implementation, the system dynamically matches safety measure entries based on the work content in the safety measure problem inspection rules, checking whether the safety measures comprehensively cover risk points and are targeted. For live-line work tasks, if the field for retained live parts is "None," or if the field for retained live parts in a meter box replacement task is "All Retained," it is marked as a safety measure problem. The system also checks whether the safety measures include voltage testing procedures, requirements for wearing safety goggles, measures to prevent short circuits or grounding in the voltage secondary circuit, and measures to prevent open circuits in the current secondary circuit. After comparing the inspection results with the knowledge base rules, the system outputs a safety measure problem mark and rectification suggestions.
[0038] This rule ensures that safety measures are highly aligned with the work content, covering all risk points and preventing operational accidents. Automated inspection mechanisms reduce human oversights, improving the safety and standardization of electrical work.
[0039] In one possible implementation, the dual-license issue checking rules combine work order information and on-site operation platform data to determine whether dual-license is required for customer-side operations. For low-voltage work orders or transformer tasks, the system automatically determines that dual-license is not required; for non-low-voltage tasks, if dual-license is not required, it is marked as a dual-license issue. The system records the issue in the knowledge base and provides correction suggestions to ensure the effective implementation of the dual-license system.
[0040] This rule ensures the strict implementation of the dual-license system in electrical operations, preventing safety hazards caused by single-license operations. Automated identification improves review efficiency and ensures the standardization of the system.
[0041] In one possible implementation, during text and visual feature extraction, the MaxKB knowledge base management system performs word segmentation, entity recognition, and relation extraction on the work order text to generate a text feature vector. The GLM-4V-Flash visual encoder, based on the ViT architecture, extracts image features, including signature areas, table structures, and handwritten text areas, generating a visual feature vector. Text and visual features are fused through a multi-level extraction pipeline to generate structured data, providing a foundation for subsequent cross-modal analysis.
[0042] The embodiments of this invention ensure that text and image information are fully extracted and fused, achieving comprehensive structuring of work order information. Cross-modal data fusion improves inference accuracy, ensuring more reliable security inspection results.
[0043] In one possible implementation, in the cross-modal attention layer, the system maps text feature vectors and visual feature vectors to a 512-dimensional space using a linear projection matrix, then inputs this to a multi-head attention mechanism to calculate the association weights of the query vector, key vector, and value vector. These association weights are used to weighted sum the text and visual feature vectors to generate a fused feature vector. During the dynamic updating of the knowledge base, new rules are added to the MaxKB knowledge base after manual review, and visual recognition error cases are annotated and added to the GLM-4V-Flash visual encoder fine-tuning dataset. The projection matrix of the cross-modal attention layer is periodically retrained to adapt to changes in feature distribution.
[0044] By employing a cross-modal attention mechanism, deep fusion of textual and visual information is achieved, improving the accuracy of work order compliance judgments. A dynamic update mechanism ensures continuous system optimization, adapting to emerging work scenarios and data distributions, and guaranteeing power operation safety and rule enforcement consistency.
[0045] Accordingly, embodiments of the present invention also provide an artificial intelligence-based work ticket security management and identification system, including a memory configured to store instructions, a processor configured to retrieve the instructions from the memory, and capable of implementing an artificial intelligence-based work ticket security management and identification method as described in any embodiment of the present invention when executing the instructions.
[0046] Example: This embodiment uses a specific emergency repair work order as an example. The work order involves emergency repair of a 10 kV low-voltage line, including a work task description, time schedule, signature area, and safety measures. The specific implementation steps of this method are as follows: Step 1: Work order data collection and format conversion; Power system staff save the completed PDF work orders to a designated directory, such as "C: / WorkOrders / ". The system uses a Python script to call the pdf2image library to convert the PDF work orders into an image format that the system can recognize (such as PNG). During the conversion process, the system preprocesses the image: Resolution standardization: The image resolution is uniformly adjusted to 300 DPI to ensure the clarity of subsequent feature extraction.
[0047] Noise Removal: The image is processed using a median filtering algorithm to eliminate salt-and-pepper noise introduced during scanning or transmission.
[0048] Image enhancement: Enhance image contrast using histogram equalization techniques to highlight text and signature areas.
[0049] For example, in a work order containing a handwritten signature and a form, the text edges in the preprocessed image are clearer, and the pixel distribution in the signature area is more uniform.
[0050] Step 2: Establishing security inspection rules for work permits; The system configures the work order safety inspection rule base through the user interface. The rule base includes the following six categories of inspection content, and each category of rules is defined based on power system safety regulations and historical accident cases: Rules for checking incorrect work order types: Determine the work order type based on the work task described on the work order. Specifically, a work order for a 35 kV voltage level is a corresponding power supply work order; a work order for a 20 kV voltage level is a corresponding distribution work order; and a work order for a 10 kV or lower voltage level is a corresponding low-voltage work order. For example, if the work order is described as "10 kV line emergency repair," but the work order type is incorrectly selected as "substation work order," the system will mark it as a misuse of work order type.
[0051] Work Task Non-Standardization Inspection Rules: Analyze the text content of the "Work Task" field in the work order. The system maintains a list of common error descriptions, including "Low-voltage New Installation / Capacity Expansion_Installation / Replacement", "Metering Equipment Replacement_Installation / Replacement", and "Account Closure_Installation / Replacement". If the text content contains the suffix "Installation / Replacement" and does not belong to the "Metering Device Failure_Installation / Replacement" category (i.e., the task content is unclear), it is judged as an unclear work task description. For example, if the work task is "Low-voltage New Installation / Capacity Expansion_Installation / Replacement", the system will mark it as a non-standard issue.
[0052] Time-related issue checking rules: Verify the logical relationships of time fields in the work order, including planned start time, planned end time, issuance time, permission time, termination time, and extension time. The system verifies the following logical chain: issuance time is earlier than planned start time, permission time is later than planned start time, termination time is earlier than planned end time, extension time is earlier than planned end time, and the work order validity period does not span multiple days (i.e., planned start and end times are on the same day). Furthermore, the system combines the number of work team members and workload to determine the reasonableness of the time arrangement. For example, if the work task is "replace 150 meters," the work team has only one member, and the planned working time is lower than a preset threshold (the preset threshold is obtained through historical data statistics; for example, if replacing one meter takes an average of 2 minutes, then replacing 150 meters would take at least 300 minutes, and the threshold is set to 300 minutes), then it is marked as a time issue.
[0053] Three signature verification rules: The system distinguishes between handwritten and machine-printed signatures by analyzing the pixel distribution of the signature area using visual features. Specifically, the system calculates the texture features of the signature area (such as a local binary pattern histogram), with handwritten signatures typically exhibiting higher texture complexity. Except for telephone authorization scenarios (where the licensor's signature can be machine-printed), all other licensor signatures must be handwritten. For example, if the licensor's signature area is detected to be in standard font and has a simple texture, it is determined to be a machine-printed signature and marked as a signature problem. The system skips the verification process if the "Work Responsible Person" and "Work Team Member" fields do not contain a signature area.
[0054] Safety measure issue inspection rules: Safety measure items are dynamically matched based on the work content. The system checks whether the safety measures are consistent with the work content, whether they comprehensively cover risk points, and whether they are targeted. For example, for live-line work tasks, if the "None" field is set to "None," it is marked as a safety measure issue; for meter box replacement tasks, if the "All Live-line Parts Retained" field is set to "All Retained," it is marked as a safety measure issue. The system also verifies whether the safety measures include necessary items, such as voltage testing procedures, requirements for wearing safety goggles, measures to prevent short circuits or grounding in the voltage secondary circuit, and measures to prevent open circuits in the current secondary circuit. These items are identified through keyword matching and semantic analysis.
[0055] Dual-permission issue checking rules: The system combines work order information with data from the on-site work platform to determine whether dual permits were required for customer-side operations. Specifically, the system extracts the work location and task from the work order and queries the permit records on the on-site work platform. If the work order is a low-voltage work order (voltage level ≤ 10 kV) or a public transformer task, dual permits are not required; otherwise, failure to execute dual permits is marked as a dual-permission issue. For example, for non-low-voltage tasks (such as 20 kV distribution work), if there are no dual-permission records on the on-site work platform, the system marks it as an issue.
[0056] Step 3: Extracting work order information and verifying key fields; The system uses the MaxKB knowledge base management system to parse the image generated in step 1 and extract key field information from the work order. The specific process includes: Text Feature Extraction: MaxKB integrates a text encoder (based on the Transformer architecture) to perform word segmentation, entity recognition, and relation extraction on the work ticket text. Word segmentation uses a dictionary-based maximum matching algorithm. Entity recognition uses a conditional random field model to identify entities such as "ticket type," "work task," and "time." Relation extraction uses a graph neural network model to construct logical relationships between entities. The text feature vector output is a tensor of dimensions [1, 128, 1024], where 128 represents the sequence length and 1024 represents the feature dimension.
[0057] Visual Feature Extraction: The GLM-4V-Flash visual encoder extracts image features based on the VisionTransformer (ViT) architecture. Specifically, the image is segmented into 16x16 pixel blocks, which are then input into the Transformer layer after linear projection and positional encoding. The visual feature vector output is a tensor of dimensions [1, 256, 1024], where 256 represents the number of image blocks and 1024 represents the feature dimension. The extracted features include signature regions (localized using the YOLOv5 object detection model), table structures (detecting lines using Hough transform), and handwritten text regions (classified using a convolutional neural network).
[0058] Multi-level information fusion: Text feature vectors and visual feature vectors are fused through a multi-level extraction pipeline. The pipeline includes an entity recognition module (outputting entity types), a relationship extraction module (outputting relationships between entities), and a rule matching module (performing preliminary matching based on the rule base from step 2). After fusion, structured data is generated, such as in JSON format, containing ticket type, work task, timestamp, signature type, security measure entries, and dual-license identifier.
[0059] Step 4: Cross-modal attention mechanism and normative identification; The system introduces a cross-modal attention mechanism to map text feature vectors and visual feature vectors to the same semantic space for joint reasoning. Feature Space Alignment: Text and visual feature vectors are projected into a 512-dimensional space using a learnable projection matrix. The projection matrix is randomly initialized and optimized during training using gradient descent. Specifically, text feature vectors are projected through a linear layer (nn.Linear(1024,512)), and visual feature vectors are projected through another linear layer (nn.Linear(1024,512)).
[0060] Attention Calculation: The projected feature vectors are input to a multi-head attention mechanism. This mechanism consists of 8 heads, each with an embedding dimension of 64 (total dimension 512). The query vector comes from the visual feature vector, while the key and value vectors come from the text feature vector. The attention weights are calculated using the Softmax function, with the formula: Attention(Q,K,V)=Softmax(QK^T / √d_k)V, where Q, K, and V represent the query, key, and value vectors, respectively, and d_k is the dimension of the key vector (64). The output is a weighted fused feature vector with dimensions [1,256,512].
[0061] Logical Reasoning: The GLM-4V-Flash logical reasoning engine performs multi-round reasoning based on fused feature vectors and the rule base from step 2. The reasoning process combines symbolic logic with neural networks. For example, for time-related issues, the engine checks whether the time field satisfies the logical constraints in the rule base; for signature-related issues, the engine combines visual features and textual descriptions to determine the signature type. The reasoning result output is a standardization judgment of the work order, including whether it passes or specifies the issue type.
[0062] Step 5: Generation and output of security assessment results; Based on the reasoning results from step 4, the system generates a work order completion standardization report. The report includes: Work tickets that meet the requirements are marked as "passed" and archived and stored in the database.
[0063] Work orders that do not meet the requirements should list the problem type, the rules and regulations violated, and the suggested corrections. For example, outputs such as "Time issue: Issuance time is later than the planned start time, it is recommended to adjust the issuance time" or "Safety measures issue: Lack of voltage testing measures, it is recommended to supplement them."
[0064] The report is displayed through a user interface and can be exported as a PDF or Excel file.
[0065] Step 6: Dynamic updating and self-optimization of the knowledge base; After each processing step, the system automatically collects newly emerging security issue cases and rule changes: New rules are added to the MaxKB knowledge base after manual review. For example, if a new accident case indicates that a certain operation requires additional safety measures, the administrator adds a rule through the backend interface, and the system updates the knowledge base through incremental learning.
[0066] Visual recognition error cases, after being labeled, were added to the fine-tuning dataset of the GLM-4V-Flash visual encoder. The labeling process included labeling the bounding boxes and categories of the error regions. The dataset was used to retrain the visual encoder, with cross-entropy loss as the loss function and Adam as the optimizer.
[0067] The system periodically retrains the projection matrix in the cross-modal attention layer to adapt to changes in feature distribution. Training data includes historical work tickets and new cases, with a training cycle of once every 30 days, and parameters are adjusted through supervised learning.
[0068] It is understood that in this embodiment: Resolution standardization: The choice of 300 DPI is based on experimental data. Values below this may result in blurry text, while values above this will increase computational overhead.
[0069] Median filtering algorithm: The filter size is 3x3, used to remove salt-and-pepper noise without destroying edge information.
[0070] Histogram equalization: Enhances the visibility of low-contrast areas by adjusting the grayscale distribution of an image.
[0071] Preset threshold (time issue): obtained through historical data statistics, for example, the average time to replace a meter is 2 minutes. The threshold is calculated as workload × unit time, and the unit time is dynamically adjusted based on the job type.
[0072] Texture feature analysis: Local binary pattern histograms are used to calculate the local texture of the signature area. The histogram entropy value of handwritten signatures is usually higher than that of machine-printed signatures.
[0073] Transformer architecture: The text encoder contains 12 Transformer layers, each with 8 headers and 1024 hidden layers.
[0074] VisionTransformer architecture: The visual encoder segments the image into 16x16 blocks and extracts features through 16 Transformer layers.
[0075] Multi-head attention mechanism: The design of 8 heads balances computational efficiency and expressive power, with each head having a dimension of 64.
[0076] Fine-tuning dataset: The labeled data includes at least 1000 samples, with a batch size of 32 and a learning rate of 0.001 during training.
[0077] The method of this invention realizes the automation, precision and self-optimization of work order safety management in power system field operations.
[0078] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.
[0079] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A work order security management and identification method based on artificial intelligence, characterized in that, The method includes the following steps: Step 1: Collect written work orders for on-site operations of the power system, convert the PDF work orders into an image format that the system can recognize, and perform resolution normalization, noise removal and image enhancement processing on the images; Step 2: Construct a work permit security inspection rule base, which includes rules for checking misuse of permit types, non-standard work tasks, time-related issues, three types of personnel and signature issues, safety measures issues, and dual-permission issues. Step 3: Parse the image generated in Step 1 using the MaxKB knowledge base management system to extract the text feature vector from the work order, and simultaneously extract the visual feature vector from the work order image using the GLM-4V-Flash visual encoder; Step 4: Input the text feature vector and visual feature vector into the cross-modal attention layer. Map the text feature vector and visual feature vector to the same semantic space through a learnable projection matrix. Calculate the association weight between the visual feature vector and the text feature vector using a multi-head attention mechanism to generate a fused feature vector. Step 5: Based on the fused feature vector, the GLM-4V-Flash logic reasoning engine is combined with the rule base of Step 2 to perform multi-round reasoning and output the normative judgment result of the work order. The judgment result includes the specific problem type of meeting or not meeting the requirements and the correction suggestions. Step 6: Based on the judgment results of Step 5, dynamically update the MaxKB knowledge base and the fine-tuning dataset of the GLM-4V-Flash visual encoder. New rules are written into the rule base after manual review, and visual recognition error cases are added to the training dataset after being labeled.
2. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The specific rules for checking ticket misuse in step 2 include: Based on the work tasks described on the work order, it is determined that the work tasks involve a 35 kV voltage level corresponding to a power supply work order, the work tasks involve a 20 kV voltage level corresponding to a power distribution work order, and the work tasks involve a 10 kV and below voltage level corresponding to a low-voltage work order. If the type of work ticket does not match the content of the work task, it will be marked as a problem of incorrect ticket type.
3. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The specific rules for checking non-standard work tasks in Step 2 include: Analyze the text content of the work task field in the work order. If the text content contains the suffix "installation, disassembly and replacement" and does not belong to the "metering device failure_installation, disassembly and replacement" type of task, it is determined that the work task description is unclear. The system maintains a list of common error descriptions, including "low-voltage new installation / capacity increase / replacement", "metering equipment replacement / replacement", and "account cancellation / replacement". The system performs real-time matching between text feature vectors and the error description list.
4. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The time-related issues checking rules in step 2 specifically include: Verify the logical relationships between the planned start time, planned end time, issuance time, permission time, termination time, and extension time in the work order, and verify that the issuance time is earlier than the planned start time, the permission time is later than the planned start time, the termination time is earlier than the planned end time, the extension time is earlier than the planned end time, and the validity period of the work order does not span days. The time arrangement is judged based on the number of team members and the workload. If the task time for a single person to replace 150 meters is less than the preset threshold, it is marked as a time problem.
5. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The specific rules for checking the three types of people and signatures in step 2 include: By analyzing the pixel distribution of the signature area using visual feature vectors, handwritten signatures can be distinguished from machine-printed signatures. Except for telephone authorization scenarios, the licensor's signature must be handwritten; if a machine-printed signature is detected, it is considered a signature problem. No judgment is needed if the "Work Supervisor" and "Work Team Member" fields do not contain a signature area.
6. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The specific rules for checking safety measures in step 2 include: Dynamically match safety measures to the work content, and check whether the safety measures are consistent with the work content, whether they fully cover the risk points, and whether they are targeted. If the field for the live parts to be retained in the live work task is "None" or the field for the live parts to be retained in the meter box replacement task is "All Retained", then it is marked as a safety issue. Verify whether the safety measures include voltage testing procedures, requirements for wearing safety goggles, measures to prevent short circuits or grounding in the voltage secondary circuit, and measures to prevent open circuits in the current secondary circuit.
7. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The specific rules for checking dual licensing issues in step 2 include: By combining the information on the work order with the data from the on-site work platform, it can be determined whether dual authorization is required for the work on the customer side. If the work order is a low-voltage work order or a transformer task, then dual authorization is not required; If a work order involves a non-low-voltage task and dual licensing is not implemented, it is marked as a dual licensing issue.
8. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, Step 3, text feature extraction and visual feature extraction, specifically includes: The MaxKB knowledge base management system performs word segmentation, entity recognition, and relation extraction on work order texts, and outputs text feature vectors. The GLM-4V-Flash visual encoder extracts image features based on the ViT architecture, including signature areas, table structures, and handwritten text areas, and outputs visual feature vectors. Textual and visual feature vectors are fused through a multi-level extraction pipeline to generate structured data.
9. The work order security management and identification method based on artificial intelligence according to claim 1, characterized in that, The cross-modal attention layer in step 4 specifically includes: Text feature vectors and visual feature vectors are mapped to a 512-dimensional space using linear projection matrices; The mapped text feature vector and visual feature vector are input into a multi-head attention mechanism to calculate the association weights between the query vector, key vector, and value vector. The association weights are used to weight and sum the text feature vectors and visual feature vectors to generate a fused feature vector. Step 6, the dynamic updating of the knowledge base, specifically includes: New rules are written into the MaxKB knowledge base after being manually reviewed; Visual recognition error cases were annotated and added to the fine-tuning dataset of the GLM-4V-Flash visual encoder; The system periodically retrains the projection matrix in the cross-modal attention layer to adapt to changes in feature distribution.
10. A work permit security management and identification system based on artificial intelligence, characterized in that, The system includes a memory configured to store instructions, a processor configured to retrieve the instructions from the memory, and, when executing the instructions, to implement the AI-based work ticket security management and identification method as described in any one of claims 1-9.