Electric power system intelligent two-ticket auxiliary decision-making method and system integrating environment perception and operation risk analysis

Abstract

The invention provides an electric power system intelligent two-ticket auxiliary decision-making method and system integrating environment perception and operation risk analysis, and relates to the technical field of electric power systems. The method comprises the following steps: acquiring environmental parameters such as temperature, humidity, wind speed, wind direction and gas concentration on a working site to form a time sequence characteristic matrix; analyzing the operation ticket and the work ticket, extracting information such as a work type, a voltage grade and personnel qualification, and generating a work feature matrix; environment and operation feature vectors are extracted through a deep network, and a joint feature vector is generated based on attention mechanism fusion; inputting the risk control model to output risk probability and grade, comparing the risk probability and grade with a rule base, identifying non-conformity items and pushing optimization suggestions; the operation state is monitored in real time, and a pause instruction and alarm are automatically sent out when a pause threshold value is reached; feedback data is collected after operation, high-value samples are screened based on a sliding time window, the model is updated in an increment or federated learning mode, and the risk identification capability is continuously improved.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a smart two-ticket auxiliary decision-making method and system for power systems that integrates environmental perception and operational risk analysis. Background Technology

[0002] In the operation and maintenance of my country's power system, operation tickets and work tickets (collectively referred to as "two tickets") are still mainly in paper or static electronic forms, completed manually, reviewed manually, and archived manually. This traditional approach relies on the experience of the operators and lacks real-time data support for the actual on-site environment, resulting in the quality of the tickets and the safety of execution being highly dependent on individual skill levels.

[0003] With the integration and application of smart grid, Internet of Things and artificial intelligence technologies, power companies are accelerating the digitalization of operation and maintenance, intelligent risk assessment and intelligent decision-making; new technologies such as multi-source sensor deployment, 5G / industrial wireless network access and federated learning are becoming important directions for improving the refinement, real-time and closed-loop management of the two-ticket system.

[0004] Despite the significant trend towards digitalization, most current two-ticket systems still suffer from the following problems: ① They cannot detect extreme environments such as high temperatures, strong winds, and lightning in real time; ② Risk analysis methods rely on empirical estimations and lack quantitative modeling of the three-element factors of "human-machine-environment"; ③ The systems themselves lack intelligent review and dynamic decision-making capabilities, resulting in lagging security and prevention measures.

[0005] Therefore, there is an urgent need for a two-vote auxiliary decision-making method that can integrate environmental perception, risk prediction and intelligent decision-making to achieve maximum safety protection and operational efficiency. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a power system intelligent two-ticket auxiliary decision-making method and system that integrates environmental perception and operational risk analysis. This method breaks through the limitations of traditional two-ticket systems that rely on static rules and manual judgment, and realizes intelligent auxiliary decision-making throughout the entire process of environmental perception, risk prediction, rule verification, and model self-evolution. This significantly improves the inherent safety level of power operations and the level of intelligence in two-ticket management.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A smart two-ticket auxiliary decision-making method for power systems that integrates environmental perception and operational risk analysis includes:

[0009] Environmental sensing units are deployed at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal image, noise and light parameters, which are then aggregated by the edge gateway to form a time-series environmental feature matrix.

[0010] The operation tickets and work tickets to be executed are parsed to extract information such as operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period and tools, and then uniformly vectorized to generate an operation feature matrix;

[0011] Deep feature extraction is performed on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vector and task feature vector. Correlation calculation is performed on the environmental feature vector and the task feature vector based on the attention mechanism and then fused to generate a joint feature vector.

[0012] The joint feature vector is input into the risk control prediction network, which outputs the risk probability and classifies the risk level according to a preset threshold.

[0013] The risk probability and risk level are correlated and reasoned with the safety rule base to identify non-compliance items such as missing safety measures, non-compliance with working conditions, abnormal equipment status, and excessive human risk. For the non-compliance items, invoice optimization suggestions are generated and pushed to the work supervisor, supervisor, and approver through at least one of mobile terminals and monitoring platforms.

[0014] During the operation, environmental parameters and operation status are monitored in real time. When the risk level reaches the preset termination threshold, a pause command is automatically triggered and an alarm is issued.

[0015] Collect job feedback data and event logs, add new samples to the continuous learning queue according to the sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning methods to improve the model's adaptive performance in different sites and new environmental scenarios.

[0016] Preferably, the bill optimization suggestions include at least the risk type, rectification measures, suggested implementation period, and priority.

[0017] Preferably, environmental sensing units are deployed at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal images, noise, and light parameters. These data are then aggregated by an edge gateway to form a time-series environmental feature matrix, including:

[0018] Based on the geographical location, voltage level, and potential risk scenarios of the work site, various types of environmental sensing units are selected; the environmental sensing units include at least: temperature sensor, humidity sensor, ultrasonic anemometer, tipping bucket rain gauge, catalytic combustible gas probe, infrared thermal imaging module, decibel meter, and illuminance meter.

[0019] The zero point and range of the environmental sensing unit were calibrated using the reference instruments specified in the national metrological verification regulations, and the traceability number was recorded.

[0020] According to the preset zoning coverage principle, the various environmental sensing units are installed on the top of the main transformer, the side of the switch cabinet, near the high-altitude insulator, the edge of the pedestrian passage and the gas leakage point. A unique network address and sampling channel number are assigned to each node of the environmental sensing unit through a 920MHz wireless self-organizing network.

[0021] The network time protocol service in the gateway is triggered by the GPS pulse second signal to synchronize the time of all nodes of the environmental sensing unit, ensuring that the time base error of each node is controlled within 10 milliseconds.

[0022] Driven by the scheduling table issued by the edge gateway, each of the environmental perception units performs data collection at a fixed sampling period, and encapsulates the single sample value with the node address and timestamp into a data frame;

[0023] Each of the environmental sensing units sends the data frames to the edge gateway via a wireless link. The edge gateway caches the arriving data frames in ascending order of timestamps and constructs an initial environmental feature matrix containing N rows and M columns according to a preset tag order.

[0024] The edge gateway invokes an anomaly detection algorithm based on three times the absolute deviation of the median to remove outliers and interpolate to complete the initial environmental feature matrix, thereby obtaining the cleaned time-series environmental feature matrix.

[0025] Preferably, the operation ticket and work order to be executed are parsed to extract the operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period, and tool information, and then uniformly vectorized and encoded to generate an operation feature matrix, including:

[0026] The system receives the input files of operation tickets and work tickets to be executed from the preset ticket management platform, and converts the input files into standardized text content through the format recognition module. The system then uses a layout structure parsing model to extract field titles, paragraph content, numbering information, and table areas to construct an initial ticket dataset containing the correspondence between field key names and text content. The input files include three types: PDF files, Microsoft Word files, and image scans.

[0027] The named entity recognition model trained based on deep learning is called on the field content of the initial dataset of the bill to extract and label the following eight categories of key semantic information: job type, voltage level, target equipment identifier, operator name, personnel qualification level, safety technical measures items, start and end time of the job and tool equipment name, forming a set of semantic tags with unique labels;

[0028] Based on the power grid operation terminology standard, equipment name standard library, unified coding rules for tools and equipment, and operation time expression dictionary, the free text extracted from the semantic tag set is subjected to standard term matching and normalization mapping, all entity items are standardized to the corresponding values ​​in the predefined vocabulary, and a set of structured field vectors is output.

[0029] For each field in the structured field vector set, vectorization encoding is performed using at least one of the preset word embedding model, one-hot encoding model, and rule template encoding method. The encoding results of all fields are then concatenated in the field order to form the job feature matrix with consistent dimensions.

[0030] Each field vector in the operation feature matrix is ​​subjected to field existence check, numerical type verification, value range comparison, and logical consistency judgment. If there are missing fields, fields that do not conform to the format, or logical conflicts, the execution step "receive the file input of the operation ticket and work ticket to be executed from the preset ticket management platform" is returned.

[0031] Preferably, deep feature extraction is performed on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vectors and task feature vectors. Then, based on an attention mechanism, correlation calculation is performed on the environmental feature vectors and the task feature vectors, and they are fused to generate a joint feature vector, including:

[0032] The zero-mean, unit-variance method is used to normalize all numerical fields of the environmental feature matrix E and the operational feature matrix W, resulting in a normalized matrix E. n With W n ;

[0033] In E respectively n With W n The system applies a one-dimensional convolutional layer with a kernel size of 3 and the GELU activation function to extract local contextual information and output a primary feature vector F. e 0 With Fw 0 ;

[0034] F e 0 With Fw 0 Each input is an independent Transformer encoder tower with two encoder layers and a fixed number of encoder heads h. After position encoding and layer normalization, the environmental feature vector F is obtained. e With the job feature vector Fw;

[0035] Within the same training round, for F e With Fw w Perform cross-attention weight calculation and generate a joint feature vector. The calculation process satisfies the following formula:

[0036]

[0037] Among them, F e F is the environment feature vector obtained by double-ended Transformer encoding; w W is the job feature vector obtained through double-ended Transformer encoding. q W k W v These are the query mapping matrix, key mapping matrix, and value mapping matrix, all automatically learned during training; d is the scaling constant for the key vector dimension, used for numerical stability; softmax is the row normalization exponential function; F f This is the final joint feature vector.

[0038] Preferably, the joint feature vector is input into the risk control prediction network, which outputs the risk probability and classifies the risk level according to a preset threshold, including:

[0039] The joint feature vector F f Input the pre-trained offline feedforward risk prediction network, perform matrix multiplication and bias addition operations between the input layer and the hidden layer to form a linear mapping result;

[0040] Applying the Sigmoid activation function to the linear mapping result yields the risk probability P, calculated as follows:

[0041]

[0042] Among them, W r b is the risk network weight matrix; r σ is the risk network bias term; P is the output risk probability; t1 is the low-to-medium risk threshold; t2 is the medium-to-high risk threshold; R is the risk level identifier obtained from the threshold determination.

[0043] An exponential moving average method is used to smooth the risk probability sequence within three consecutive time windows to improve the robustness of sudden data. Then, the smoothed probability value is compared with the level R of the previous state. If a jump across two levels occurs, a model self-check signal is triggered.

[0044] Preferably, the risk probability and risk level are correlated with the safety rule base to determine non-compliance items such as missing safety measures, unsuitable working conditions, abnormal equipment status, and excessive human-caused risks. For these non-compliance items, invoice optimization suggestions are generated and pushed to the work supervisor, supervisor, and approver via at least one of mobile terminals or a monitoring platform, including:

[0045] Load rule entries that match the current job type, voltage level, equipment name, and risk level, and compare the job characteristics with the rule entries one by one to identify non-compliance items such as missing safety technical measures, non-compliance with job conditions, abnormal equipment status, and insufficient personnel qualifications, and classify and mark the non-compliance items according to risk type.

[0046] Based on the risk type and severity level of each non-compliance, generate invoice optimization suggestions that include rectification recommendations, implementation time periods, and priorities, and transmit them to the mobile terminals configured by the work supervisor, work monitor, and invoice approver.

[0047] The optimization suggestions are pushed out in the form of structured messages, and the sending status and user response results are recorded.

[0048] Preferably, environmental parameters and operation status are monitored in real time during the operation. When the risk level reaches a preset termination threshold, a pause command is automatically triggered and an alarm is issued, including:

[0049] The system acquires real-time data on personnel entry and exit, tool usage, operation step confirmation, and equipment start / stop status during the operation process from on-site operation recording terminals, equipment worn by operators, and tool barcode recognition devices, and simultaneously constructs a set of operation execution statuses.

[0050] Temperature, wind speed, combustible gas concentration, lightning warning value, and noise level at the current work site are collected at fixed time intervals. The temperature, wind speed, combustible gas concentration, lightning warning value, and noise level at the current work site are then combined with the work execution status set and uploaded to the edge computing node as risk assessment input for the current monitoring period.

[0051] Call the deployed risk control prediction model, combine the current environmental parameters and operation execution status, determine whether a new risk level result is generated, and compare it with the preset termination threshold.

[0052] When the current risk level reaches or exceeds the suspension threshold, a suspension command is sent to the operation terminal via the local communication bus, and multiple suspension alarm messages are simultaneously issued through the on-site audible and visual alarm, the alarm window of the monitoring platform, and the mobile terminal of the operator.

[0053] Preferably, the risk control prediction network is updated by collecting operation feedback data and event logs, adding new samples to the continuous learning queue according to a sliding time window, and using at least one of incremental learning and federated learning methods, including:

[0054] The system automatically collects data on the work process after the task is completed. This data includes operator logs, environmental parameter change records, system alarm records, invoice approval process logs, task completion confirmation information, and post-task evaluation records.

[0055] The operation process data is archived and integrated according to the operation number to form a complete operation feedback sample; samples that have experienced increased risk, violations, rule non-compliance, or alarm response events during the operation are marked as high-value samples and added to the training sample pool.

[0056] According to the sample update strategy of the preset sliding time window, recent data is selected from the sample pool to be trained and added to the continuous learning queue, and expired samples that exceed the time window range are deleted to ensure that the training data is always fresh and conforms to the characteristics of the current operation scenario.

[0057] The model incremental training task is executed in the central server or regional edge node. Based on the current continuous learning queue, the risk control prediction model weights are updated through incremental learning or federated learning, and model version management and deployment synchronization operations are completed.

[0058] A smart two-ticket auxiliary decision-making method for power systems that integrates environmental perception and operational risk analysis includes:

[0059] The environmental sensing unit is used to deploy environmental sensing units at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal image, noise and light parameters, and then aggregate them through the edge gateway to form a time-series environmental feature matrix.

[0060] The ticket information parsing unit is used to parse the operation tickets and work tickets to be executed, extract the operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period and tool information, and generate an operation feature matrix by uniform vectorization encoding;

[0061] The multimodal feature fusion unit is used to perform deep feature extraction on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vector and task feature vector, and to perform correlation calculation and fusion on the environmental feature vector and the task feature vector based on the attention mechanism to generate a joint feature vector;

[0062] The risk prediction unit is used to input the joint feature vector into the risk control prediction network, output the risk probability, and classify the risk level according to the preset threshold.

[0063] The rule verification and suggestion generation unit is used to correlate and reason with the risk probability and risk level with the safety rule base to determine non-compliance items such as missing safety measures, non-compliance with working conditions, abnormal equipment status, and excessive human risk. It generates invoice optimization suggestions for the non-compliance items and pushes the invoice optimization suggestions to the work supervisor, supervisor, and approver through at least one of the following methods: mobile terminal and monitoring platform.

[0064] The operation implementation monitoring and termination control unit is used to monitor environmental parameters and operation status in real time during operation implementation. When the risk level reaches the preset termination threshold, it automatically triggers a pause command and issues an alarm.

[0065] The model adaptive update unit is used to collect job feedback data and event logs, add new samples to the continuous learning queue according to the sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning methods to improve the model's adaptive performance in different sites and new environmental scenarios.

[0066] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0067] Traditional two-ticket systems cannot detect changes in the external environment such as high temperatures, lightning, and strong winds in real time, posing a safety hazard of "disconnect between operational information and actual conditions." This invention constructs an on-site environmental sensing unit to dynamically collect key parameters such as temperature, humidity, wind speed, and gas concentration, and integrates them into the decision-making process. This enables comprehensive perception of the external environment before ticket approval, improving the adaptability and timeliness of the two-ticket system from the source.

[0068] Most existing methods rely on static rule-based judgments, which cannot handle the complex risk coupling caused by the interaction of multiple sources of information. This invention introduces a deep neural network model to fuse and model operational information and environmental data, and outputs risk probabilities and levels, breaking through the bottleneck of experience-based judgment modes and realizing quantitative assessment and dynamic response to operational safety risks.

[0069] Traditional two-ticket approval processes rely on human experience, making it difficult to detect deficiencies in safety measures or abnormal staffing. This invention uses a safety rule base and model output results to jointly perform associative reasoning, intelligently identifying key non-compliance items in the operation and automatically generating rectification suggestions, which are then pushed to the responsible person and approver, significantly improving the accuracy and standardization of two-ticket approval.

[0070] Existing systems struggle to adapt to environmental differences and risk variations across different work sites, and their models are prone to aging. This invention introduces a sliding time window, continuous sample collection, and incremental / federated learning mechanisms to enable the risk control model to continuously evolve, effectively improving its adaptability and accuracy under different operational conditions. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 A flowchart of the method provided in an embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the system structure provided in an embodiment of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] The purpose of this invention is to provide a smart two-ticket auxiliary decision-making method and system for power systems that integrates environmental perception and operational risk analysis. It breaks through the limitations of traditional two-ticket systems that rely on static rules and manual judgment, and realizes intelligent auxiliary decision-making throughout the entire process of environmental perception, risk prediction, rule verification and model self-evolution, which greatly improves the inherent safety level of power operations and the level of intelligence of two-ticket management.

[0076] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0077] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a smart two-ticket auxiliary decision-making method for power systems that integrates environmental perception and operational risk analysis, comprising:

[0078] Step 100: Deploy environmental sensing units at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal image, noise and light parameters, and then aggregate them through the edge gateway to form a time-series environmental feature matrix.

[0079] Step 200: Parse the operation ticket and work order to be executed, extract the operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period and tool information, and generate an operation feature matrix by uniform vectorization encoding;

[0080] Step 300: Perform deep feature extraction on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vector and task feature vector. Then, perform correlation calculation and fusion on the environmental feature vector and task feature vector based on the attention mechanism to generate a joint feature vector.

[0081] Step 400: Input the joint feature vector into the risk control prediction network, output the risk probability, and classify the risk level according to the preset threshold;

[0082] Step 500: Link the risk probability and risk level with the safety rule base to identify non-compliance items such as missing safety measures, non-compliant working conditions, abnormal equipment status, and excessive human-caused risks. Generate invoice optimization suggestions for non-compliance items and push the invoice optimization suggestions to the work supervisor, supervisor, and approver through at least one of the following methods: mobile terminal, monitoring platform. The invoice optimization suggestions shall include at least the risk type, rectification measures, suggested execution period, and priority.

[0083] Step 600: Monitor environmental parameters and operation status in real time during the operation. When the risk level reaches the preset termination threshold, automatically trigger a pause command and issue an alarm.

[0084] Step 700: Collect job feedback data and event logs, add new samples to the continuous learning queue according to the sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning to improve the model's adaptive performance in different sites and new environmental scenarios.

[0085] In step 100 of this embodiment, it is preferable to deploy multiple types of environmental sensing units at the power operation site to perceive environmental parameters related to operational safety in real time. Specifically, sensors are first selected based on the geographical location, voltage level, equipment type, and potential risk factors (such as lightning-prone areas, high-voltage testing environments, and oil and gas diffusion risk areas) of the target operation area. The selected environmental sensing units include at least: temperature sensors, humidity sensors, ultrasonic anemometers, tipping bucket rain gauges, catalytic combustible gas probes, infrared thermal imaging modules, decibel meters, and illuminance meters. To ensure measurement accuracy, all sensors undergo zero-point calibration and range calibration using standard equipment specified in national metrological verification regulations, and corresponding traceability numbers are recorded for later traceability.

[0086] Furthermore, the environmental sensing units are deployed in key power equipment and personnel access areas according to the principle of "functional zoning + risk sensitivity priority," including the top of the main transformer, the side of the switchgear, near high-altitude insulators, the edge of pedestrian passages, and low-lying areas of the site where flammable gas leaks may occur. All nodes are connected to a 920MHz frequency band wireless self-organizing network communication protocol, automatically forming a network, and each node is uniformly assigned a unique network address and corresponding sampling channel number by the edge gateway. For time synchronization, the system uses a GPS pulse-second signal to trigger the network time protocol module built into the edge gateway to perform millisecond-level time synchronization of all environmental sensing nodes in the site, ensuring that the sampling time deviation between nodes is controlled within 10 milliseconds, meeting the high synchronization accuracy requirements.

[0087] Furthermore, each environmental sensing unit collects data once every preset sampling period (e.g., every 30 seconds) and sends structured data frames containing node addresses, measured values, and timestamps to the edge gateway via a wireless link. The edge gateway arranges the data frames in ascending order of timestamps and constructs an initial environmental feature matrix based on sensor type labels and sampling channel numbers. The matrix uses rows to represent time steps and columns to represent various environmental parameters. To improve data quality, the gateway internally calls an algorithm based on "three times the absolute deviation of the median" to perform anomaly detection and removal on parameter values ​​within each time step, and performs linear interpolation completion processing on data points identified as outliers. Finally, a stable, complete, and time-series-characteristic-rich cleaned environmental feature matrix is ​​generated for use by subsequent risk analysis models.

[0088] In step 200 of this embodiment, it is preferable to receive operation tickets and work tickets to be executed from a preset ticket management platform. The input file formats include PDF files, Microsoft Word files, and image scans. This embodiment first performs format conversion processing on the received ticket files. Image-type tickets are converted into text content using OCR text recognition, while other files directly extract text paragraphs and table structures. Subsequently, based on a layout structure analysis model, the text content is partitioned and identified, extracting field titles, paragraph content, number identifiers, and table areas to construct an initial dataset of tickets with key-value correspondence, which serves as input data for subsequent semantic analysis.

[0089] Furthermore, this embodiment performs semantic extraction on the text fields in the initial dataset based on a pre-trained named entity recognition model. This embodiment can identify eight key semantic information categories, specifically including: job type, voltage level, target equipment identifier, operator name, personnel qualification level, safety technical measures, job start and end times, and tool / equipment name. The recognition results are represented as a set of semantic tags, with each tag bound to a unique identifier and corresponding text content. Subsequently, this embodiment calls the power grid operation terminology standard, equipment standard naming table, unified tool coding dictionary, and operation time expression dictionary to standardize and match the text items in the semantic tag set, outputting a structured field vector set, providing a unified semantic foundation for subsequent encoding.

[0090] Furthermore, this embodiment employs word embedding models, one-hot encoding, or rule templates to quantify the content of each field in the structured field vector set, based on field type and modeling requirements. All encoded field vectors are concatenated in a preset order to form a consistent-dimensional job feature matrix. This embodiment further performs integrity checks on this feature matrix, specifically including field existence checks, format validity checks, value range comparisons, and logical consistency verification between fields. If missing fields, abnormal data formats, or logical conflicts exist between job content and personnel qualifications or voltage levels, the file extraction step is automatically re-executed until all checks pass before proceeding to the feature fusion and risk analysis process.

[0091] In step 300 of this embodiment, the environmental feature matrix formed in step 100 and the job feature matrix formed in step 200 are first preprocessed to improve the numerical stability and convergence efficiency in the subsequent modeling process. Specifically, this embodiment uses a zero-mean, unit-variance normalization method to calculate the mean and standard deviation of all numerical fields in the two feature matrices column-wise, and performs standardization transformation to generate normalized feature matrices. Subsequently, a one-dimensional convolutional neural network structure is applied to the normalization results, with the kernel size set to 3, the stride to 1, and the number of channels set between 16 and 64 depending on the input dimension. A GELU activation function is then applied after each convolutional layer to extract local contextual change patterns and output the initially extracted feature vector as the basic representation for deep fusion.

[0092] To further extract long-term dependency information and global features, this embodiment inputs the aforementioned primary feature vectors into two structurally symmetrical Transformer encoder modules. Each encoder consists of two stacked layers, internally containing a multi-head attention mechanism, a feedforward network, and a layer normalization module. The number of heads in the multi-head attention mechanism is fixed at 8, and the dimension of the attention hidden layer is set to range from 128 to 256. Sine and cosine positional encoding is introduced during the encoding process to preserve temporal information. After the Transformer encoding is completed, environmental feature vectors and job feature vectors are obtained respectively. Subsequently, a cross-attention mechanism is executed in the fusion stage to calculate the correlation between the representation spaces of the two feature vectors and establish a fusion channel, finally generating a fused joint feature vector, providing a unified input for subsequent risk prediction.

[0093] Specifically, all network parameters involved in this embodiment (including the number of convolutional kernels, the number of Transformer layers, the number of multi-head attention heads, the encoding dimension, and the normalization method) were evaluated and hyperparameters tuned on the training set through prior experiments. The convolutional kernel size was uniformly set to 3, the number of multi-head attention heads ranged from 4 to 8, and the representation dimension of the attention network was controlled between 128 and 256 to achieve a balance between model complexity and computational resources. These parameters were selected using a grid search strategy on historical job data from multiple sites and were consistently applied to the model deployment version. The weight matrix and bias terms involved in the attention mechanism were automatically learned through backpropagation during training, requiring no manual specification, ensuring the consistency and reproducibility of the parameter convergence process.

[0094] In step 400 of this embodiment, the joint feature vector generated in step 300 is used as input to the feedforward risk prediction network that has completed offline training. This network consists of an input layer, at least one hidden layer, and an output layer. The hidden layer adopts a fully connected structure, with the number of nodes set between 64 and 128. Linear mapping operations are performed between each layer, and a Sigmoid activation function is applied to the output to compress the prediction result to a probability value between 0 and 1, which serves as the risk probability of the task. To enhance the stability against fluctuations in task status, this embodiment further performs exponential moving average processing on the risk probability results over three consecutive time periods. If the current smoothed risk probability, compared to the risk level defined in the previous period, jumps directly from low risk to high risk or from high risk to low risk—that is, a sudden change in level exceeding one level—the risk model's self-checking mechanism is immediately triggered to determine the rationality of the model's output.

[0095] In this embodiment, the probability thresholds used to classify risk levels are pre-set and optimized through large-sample testing. The preferred boundary between low and medium risk is between 0.3 and 0.4, and the preferred boundary between medium and high risk is between 0.6 and 0.7. Specific values ​​are determined based on the actual distribution density of historical operation risk labels. All weight matrices and bias parameters in the network structure are obtained offline through standard backpropagation algorithm on historical labeled datasets and are frozen to the deployment version when the convergence condition is met, eliminating the need for manual setting after deployment. The time window length for the moving average is set to 3 sampling periods, and the smoothing coefficient is set between 0.3 and 0.5 to suppress the impact of occasional anomalies while ensuring response speed. These parameters were obtained through grid search and cross-testing on the validation set in pre-experiments, demonstrating good generalization performance and practical stability.

[0096] In step 500 of this embodiment, after determining the risk probability and risk level, a locally deployed safety rule base is preferably invoked. This rule base is indexed in multiple dimensions according to job type, voltage level, equipment type, and risk level. Rule entries include risk source descriptions, judgment logic conditions, safety technical measure templates, and personnel qualification constraints. For the current job to be executed, the system sequentially loads rule entries matching the job characteristics and compares the field values ​​in the job feature matrix with the rule logic conditions one-to-one. If any job attribute does not match a rule entry, it is marked as a non-compliance item. Identified problems are categorized according to risk source type, including four categories: missing safety measures, non-compliant job conditions, abnormal equipment status, and insufficient personnel qualifications, to support accurate matching in subsequent suggestion generation.

[0097] Optionally, based on the risk source category and severity level of each non-compliance item, this embodiment preferably retrieves a structured suggestion module from the suggestion generation template library to automatically generate invoice optimization suggestions. The suggestions include at least: suggested corrective measures (such as adding temporary grounding, safety fencing, or personnel replacement), suggested execution time (e.g., avoiding the midday heat or thunderstorm warning periods), and priority level (urgent, medium, and deferred). After generation, this embodiment, based on the terminal configuration of the work supervisor, work monitor, and invoice approver, selects one or more channels from mobile terminal application push or monitoring platform interface display to deliver the suggestion information in the form of a structured message and record the receipt status, thus achieving a closed-loop notification.

[0098] Specifically, during the formal execution phase of the operation, this embodiment uses a monitoring cycle of 60 seconds to collect environmental data and operation process data at the current work site. Environmental data includes temperature, wind speed, combustible gas concentration, lightning warning values, and noise levels, sourced from the environmental sensing unit described in step 100. Operation process data comes from personnel-worn terminals, tool barcode readers, and operation recording terminals, including personnel entry / exit records, tool usage records, confirmation results of key operation steps, and equipment status change logs. This embodiment encapsulates the above data into a dynamic status input for the current monitoring cycle, which is transmitted to the edge computing node via an edge gateway for real-time risk assessment.

[0099] Furthermore, within each monitoring cycle, this embodiment invokes the deployed risk control prediction model, using the input environmental parameters and operation execution status as feature inputs, and outputs the current risk level. When the current level reaches or exceeds a preset termination threshold (e.g., set to medium to high level or above), this embodiment immediately issues a pause command to the operation console via the edge control terminal and simultaneously activates the audible and visual alarm device. Based on this, the system will also push termination alarm information through the monitoring platform's alarm window and the operator's mobile terminal, ensuring multi-channel synchronous notification to relevant on-site personnel to immediately stop risky operations and await instructions for reassessment before resuming operation, thus forming a comprehensive high-risk linkage control mechanism.

[0100] Furthermore, in step 600 of this embodiment, during the operation implementation phase, environmental parameters and operation execution status are preferably monitored in real time to support the triggering judgment of the high-risk dynamic termination mechanism. First, this embodiment uses on-site operation recording terminals, smart positioning tags worn by workers, and barcode scanning and identification devices set at tool entry and exit control points to collect records of personnel entry and exit, tool issuance and return data, execution confirmation results of key operation steps, and changes in equipment start / stop status. The operation execution status data is synchronously aggregated with a 10-second refresh cycle to construct a real-time operation status set.

[0101] Based on this, this embodiment collects environmental parameters of the current work site every 60 seconds. These environmental parameters include: temperature (range -10℃ to 60℃, accuracy not less than ±0.5℃), wind speed (range 0 to 25 m / s, accuracy not less than ±0.3 m / s), combustible gas concentration (0% to 100% LEL, response time less than 30 seconds), lightning warning level (based on national warning standards 0-3), and noise level (30 dB to 120 dB, accuracy ±1.5 dB). These environmental parameters and the work status set are packaged together in each sampling cycle and uploaded to the edge computing device deployed on-site via the edge acquisition node, serving as input data for the current monitoring cycle.

[0102] The risk control prediction model is pre-loaded into the edge device and performs a risk assessment task once per monitoring cycle. Based on historical training weights and the input data of the current cycle, the model outputs the latest risk level, which is set according to a five-level standard: extremely low, relatively low, moderate, relatively high, and extremely high. The preset termination threshold is preferably set to "relatively high" or above; that is, when the model outputs a level of "relatively high" or "extremely high," the current operation is considered to be in a dangerous state. Once the current risk level is determined to reach or exceed the termination threshold, this embodiment immediately issues a stop operation command to the operating terminal via the RS485 communication bus and simultaneously activates the audible and visual alarms located in the work area (alarm response time no greater than 1 second). Simultaneously, a termination alarm is sent wirelessly to the operator's mobile terminal and the alarm module of the remote monitoring platform, realizing a multi-channel parallel notification mechanism. The termination status is marked as a lock screen on the operating terminal, requiring review and authorization by authorized personnel with specific permissions before the operation can resume, thereby ensuring immediate control and closed-loop handling of medium- to high-risk operations.

[0103] In step 700 of this embodiment, after each work process is completed, six types of key data related to the current work process are automatically collected, including operator operation logs, records of changes in environmental parameters at the work site, system alarm records triggered during the work, approval process logs for operation tickets and work tickets, work completion confirmation data, and post-work evaluation form content. The above data is archived and organized by work number, forming a structured work feedback sample set. This embodiment filters samples based on whether high-risk events such as significantly increased risk level, operational violations, rule matching failures, or alarm response handling occur during work execution. If such events occur, the sample is marked as a high-value sample and added to the training sample pool for subsequent model updates.

[0104] To ensure the timeliness and representativeness of the model input data, this embodiment employs a sliding time window strategy to manage sample data flow. Specifically, before each model update, high-value samples from a past period (e.g., the last 7 days or the last 100 jobs) are extracted from the sample pool and added to the continuous learning queue. Simultaneously, the system automatically removes expired samples outside the time window to prevent old data from interfering with the model's ability to judge the current scenario. The length and update frequency of the sliding window can be adjusted according to the actual deployment environment; it is recommended to set it to a periodic configuration of 24 hours, 3 days, or 7 days to balance data coverage breadth and model response speed. Through this strategy, the continuous learning queue always maintains the "freshness" and contextual coherence of the data, providing a stable data input source for incremental training.

[0105] This embodiment automatically initiates a model training task on the central server or regional edge nodes after each learning queue update, updating the weight parameters of the original risk control prediction model using either incremental learning or federated learning. Incremental learning is suitable for data accumulation at a single operational site, while federated learning supports distributed model collaborative updates across multiple operational sites, ensuring data privacy while improving model generalization ability. The updated model version is uniformly numbered and archived by the model management module, and the new version is deployed to the inference engine used on-site. The model synchronization mechanism supports hot loading and canary updates to ensure uninterrupted operation, and has a rollback mechanism to restore to the previous stable version in case of model performance anomalies, thereby achieving continuous model evolution and reliable operation and maintenance.

[0106] Corresponding to the above methods, such as Figure 2 As shown in the figure, this embodiment also provides a smart two-ticket auxiliary decision-making method for power systems that integrates environmental perception and operational risk analysis, including:

[0107] The environmental sensing unit is used to deploy environmental sensing units at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal image, noise and light parameters, and then aggregate them through the edge gateway to form a time-series environmental feature matrix.

[0108] The ticket information parsing unit is used to parse the operation tickets and work tickets to be executed, extract the operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period and tool information, and generate an operation feature matrix by uniform vectorization encoding;

[0109] The multimodal feature fusion unit is used to perform deep feature extraction on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vector and task feature vector, and to perform correlation calculation and fusion on the environmental feature vector and the task feature vector based on the attention mechanism to generate a joint feature vector;

[0110] The risk prediction unit is used to input the joint feature vector into the risk control prediction network, output the risk probability, and classify the risk level according to the preset threshold.

[0111] The rule verification and suggestion generation unit is used to correlate and reason with the risk probability and risk level with the safety rule base to determine non-compliance items such as missing safety measures, non-compliance with working conditions, abnormal equipment status, and excessive human risk. It generates invoice optimization suggestions for the non-compliance items and pushes the invoice optimization suggestions to the work supervisor, supervisor, and approver through at least one of the following methods: mobile terminal and monitoring platform.

[0112] The operation implementation monitoring and termination control unit is used to monitor environmental parameters and operation status in real time during operation implementation. When the risk level reaches the preset termination threshold, it automatically triggers a pause command and issues an alarm.

[0113] The model adaptive update unit is used to collect job feedback data and event logs, add new samples to the continuous learning queue according to the sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning methods to improve the model's adaptive performance in different sites and new environmental scenarios.

[0114] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0115] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A smart two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis, characterized in that, include: Environmental sensing units are deployed at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal image, noise and light parameters, which are then aggregated by the edge gateway to form a time-series environmental feature matrix. The operation tickets and work tickets to be executed are parsed to extract information such as operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period and tools, and then uniformly vectorized to generate an operation feature matrix; Deep feature extraction is performed on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vector and task feature vector. Correlation calculation is performed on the environmental feature vector and the task feature vector based on the attention mechanism and then fused to generate a joint feature vector. The joint feature vector is input into the risk control prediction network, which outputs the risk probability and classifies the risk level according to a preset threshold. The risk probability and risk level are correlated and reasoned with the safety rule base to identify non-compliance items such as missing safety measures, non-compliance with working conditions, abnormal equipment status, and excessive human risk. For the non-compliance items, invoice optimization suggestions are generated and pushed to the work supervisor, supervisor, and approver through at least one of mobile terminals and monitoring platforms. During the operation, environmental parameters and operation status are monitored in real time. When the risk level reaches the preset termination threshold, a pause command is automatically triggered and an alarm is issued. Collect job feedback data and event logs, add new samples to the continuous learning queue according to the sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning methods to improve the model's adaptive performance in different sites and new environmental scenarios.

2. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, The proposed bill optimization suggestions should include at least the risk type, rectification measures, suggested implementation period, and priority.

3. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, Environmental sensing units are deployed at the work site to periodically collect data on temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal images, noise, and illumination parameters. These data are then aggregated by an edge gateway to form a temporal environmental feature matrix, including: Based on the geographical location, voltage level, and potential risk scenarios of the work site, various types of environmental sensing units are selected; the environmental sensing units include at least: temperature sensor, humidity sensor, ultrasonic anemometer, tipping bucket rain gauge, catalytic combustible gas probe, infrared thermal imaging module, decibel meter, and illuminance meter. The zero point and range of the environmental sensing unit were calibrated using the reference instruments specified in the national metrological verification regulations, and the traceability number was recorded. According to the preset zoning coverage principle, the various environmental sensing units are installed on the top of the main transformer, the side of the switch cabinet, near the high-altitude insulator, the edge of the pedestrian passage and the gas leakage point. A unique network address and sampling channel number are assigned to each node of the environmental sensing unit through a 920MHz wireless self-organizing network. The network time protocol service in the gateway is triggered by the GPS pulse second signal to synchronize the time of all nodes of the environmental sensing unit, ensuring that the time base error of each node is controlled within 10 milliseconds. Driven by the scheduling table issued by the edge gateway, each of the environmental perception units performs data collection at a fixed sampling period, and encapsulates the single sample value with the node address and timestamp into a data frame; Each of the environmental sensing units sends the data frames to the edge gateway via a wireless link. The edge gateway caches the arriving data frames in ascending order of timestamps and constructs an initial environmental feature matrix containing N rows and M columns according to a preset tag order. The edge gateway invokes an anomaly detection algorithm based on three times the absolute deviation of the median to remove outliers and interpolate to complete the initial environmental feature matrix, thereby obtaining the cleaned time-series environmental feature matrix.

4. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, The operation tickets and work orders to be executed are parsed to extract information such as job type, voltage level, target equipment, operator qualifications, safety measures, planned time period, and tools. A unified vectorized encoding is then used to generate a job feature matrix, including: The system receives the input files of operation tickets and work tickets to be executed from the preset ticket management platform, and converts the input files into standardized text content through the format recognition module. The system then uses a layout structure parsing model to extract field titles, paragraph content, numbering information, and table areas to construct an initial ticket dataset containing the correspondence between field key names and text content. The input files include three types: PDF files, Microsoft Word files, and image scans. The named entity recognition model trained based on deep learning is called on the field content of the initial dataset of the bill to extract and label the following eight categories of key semantic information: job type, voltage level, target equipment identifier, operator name, personnel qualification level, safety technical measures items, start and end time of the job and tool equipment name, forming a set of semantic tags with unique labels; Based on the power grid operation terminology standard, equipment name standard library, unified coding rules for tools and equipment, and operation time expression dictionary, the free text extracted from the semantic tag set is subjected to standard term matching and normalization mapping, all entity items are standardized to the corresponding values ​​in the predefined vocabulary, and a set of structured field vectors is output. For each field in the structured field vector set, vectorization encoding is performed using at least one of the preset word embedding model, one-hot encoding model, and rule template encoding method. The encoding results of all fields are then concatenated in the field order to form the job feature matrix with consistent dimensions. Each field vector in the job feature matrix is ​​checked for existence, value type, value range, and logical consistency. If there are missing fields, incorrect field formats, or logical conflicts, the process returns to the step "Receive the file input of the operation ticket and work ticket to be executed from the preset ticket management platform".

5. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, Deep feature extraction is performed on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vectors and task feature vectors. Then, based on an attention mechanism, correlation calculation is performed on the environmental feature vectors and the task feature vectors, and they are fused to generate a joint feature vector, including: The zero-mean, unit-variance method is used to normalize all numerical fields of the environmental feature matrix E and the operational feature matrix W, resulting in a normalized matrix E. n With W n ; In E respectively n With W n The system applies a one-dimensional convolutional layer with a kernel size of 3 and the GELU activation function to extract local contextual information and output a primary feature vector F. e 0 With Fw 0 ; F e 0 With Fw 0 Each input is an independent Transformer encoder tower with two encoder layers and a fixed number of encoder heads h. After position encoding and layer normalization, the environmental feature vector F is obtained. e With the job feature vector Fw; Within the same training round, for F e With Fw w Perform cross-attention weight calculation and generate a joint feature vector. The calculation process satisfies the following formula: Among them, F e F is the environment feature vector obtained by double-ended Transformer encoding; w W is the job feature vector obtained through double-ended Transformer encoding. q W k W v These are the query mapping matrix, key mapping matrix, and value mapping matrix, all automatically learned during training; d is the scaling constant for the key vector dimension, used for numerical stability; softmax is the row normalization exponential function; F f This is the final joint feature vector.

6. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, The joint feature vector is input into the risk control prediction network, which outputs the risk probability and classifies the risk level according to a preset threshold, including: The joint feature vector F f Input the pre-trained offline feedforward risk prediction network, perform matrix multiplication and bias addition operations between the input layer and the hidden layer to form a linear mapping result; Applying the Sigmoid activation function to the linear mapping result yields the risk probability P, calculated as follows: Among them, W r b is the risk network weight matrix; r σ is the risk network bias term; P is the output risk probability; t1 is the low-to-medium risk threshold; t2 is the medium-to-high risk threshold; R is the risk level identifier obtained from the threshold determination. An exponential moving average method is used to smooth the risk probability sequence within three consecutive time windows to improve the robustness of sudden data. Then, the smoothed probability value is compared with the level R of the previous state. If a jump across two levels occurs, a model self-check signal is triggered.

7. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, The risk probability and risk level are correlated with the safety rule base to determine non-compliance items such as missing safety measures, unsuitable working conditions, abnormal equipment status, and excessive human-caused risks. For these non-compliance items, invoice optimization suggestions are generated and pushed to the work supervisor, supervisor, and approver via at least one of the following methods: mobile terminal, monitoring platform. Load rule entries that match the current job type, voltage level, equipment name, and risk level, and compare the job characteristics with the rule entries one by one to identify non-compliance items such as missing safety technical measures, non-compliance with job conditions, abnormal equipment status, and insufficient personnel qualifications, and classify and mark the non-compliance items according to risk type. Based on the risk type and severity level of each non-compliance, generate invoice optimization suggestions that include rectification recommendations, implementation time periods, and priorities, and transmit them to the mobile terminals configured by the work supervisor, work monitor, and invoice approver. The optimization suggestions are pushed out in the form of structured messages, and the sending status and user response results are recorded.

8. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis according to claim 1, characterized in that, During the operation, environmental parameters and operation status are monitored in real time. When the risk level reaches a preset termination threshold, a pause command is automatically triggered and an alarm is issued, including: The system acquires real-time data on personnel entry and exit, tool usage, operation step confirmation, and equipment start / stop status during the operation process from on-site operation recording terminals, equipment worn by operators, and tool barcode recognition devices, and simultaneously constructs a set of operation execution statuses. Temperature, wind speed, combustible gas concentration, lightning warning value, and noise level at the current work site are collected at fixed time intervals. The temperature, wind speed, combustible gas concentration, lightning warning value, and noise level at the current work site are then combined with the work execution status set and uploaded to the edge computing node as risk assessment input for the current monitoring period. Call the deployed risk control prediction model, combine the current environmental parameters and operation execution status, determine whether a new risk level result is generated, and compare it with the preset termination threshold. When the current risk level reaches or exceeds the suspension threshold, a suspension command is sent to the operation terminal via the local communication bus, and multiple suspension alarm messages are simultaneously issued through the on-site audible and visual alarm, the alarm window of the monitoring platform, and the mobile terminal of the operator.

9. The intelligent two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis as described in claim 1, characterized in that, Collect job feedback data and event logs, add new samples to the continuous learning queue according to a sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning methods, including: The system automatically collects data on the work process after the task is completed. This data includes operator logs, environmental parameter change records, system alarm records, invoice approval process logs, task completion confirmation information, and post-task evaluation records. The operation process data is archived and integrated according to the operation number to form a complete operation feedback sample; samples that have experienced increased risk, violations, rule non-compliance, or alarm response events during the operation are marked as high-value samples and added to the training sample pool. According to the sample update strategy of the preset sliding time window, recent data is selected from the sample pool to be trained and added to the continuous learning queue, and expired samples that exceed the time window range are deleted to ensure that the training data is always fresh and conforms to the characteristics of the current operation scenario. The model incremental training task is executed in the central server or regional edge node. Based on the current continuous learning queue, the risk control prediction model weights are updated through incremental learning or federated learning, and model version management and deployment synchronization operations are completed.

10. A smart two-ticket auxiliary decision-making method for power systems integrating environmental perception and operational risk analysis, characterized in that, include: The environmental sensing unit is used to deploy environmental sensing units at the work site to periodically collect temperature, humidity, wind speed, wind direction, rainfall, combustible gas concentration, infrared thermal image, noise and light parameters, and then aggregate them through the edge gateway to form a time-series environmental feature matrix. The ticket information parsing unit is used to parse the operation tickets and work tickets to be executed, extract the operation type, voltage level, target equipment, operator qualifications, safety measures, planned time period and tool information, and generate an operation feature matrix by uniform vectorization encoding; The multimodal feature fusion unit is used to perform deep feature extraction on the environmental feature matrix and the task feature matrix respectively to obtain environmental feature vector and task feature vector, and to perform correlation calculation and fusion on the environmental feature vector and the task feature vector based on the attention mechanism to generate a joint feature vector; The risk prediction unit is used to input the joint feature vector into the risk control prediction network, output the risk probability, and classify the risk level according to the preset threshold. The rule verification and suggestion generation unit is used to correlate and reason with the risk probability and risk level with the safety rule base to determine non-compliance items such as missing safety measures, non-compliance with working conditions, abnormal equipment status, and excessive human risk. It generates invoice optimization suggestions for the non-compliance items and pushes the invoice optimization suggestions to the work supervisor, supervisor, and approver through at least one of the following methods: mobile terminal and monitoring platform. The operation implementation monitoring and termination control unit is used to monitor environmental parameters and operation status in real time during operation implementation. When the risk level reaches the preset termination threshold, it automatically triggers a pause command and issues an alarm. The model adaptive update unit is used to collect job feedback data and event logs, add new samples to the continuous learning queue according to the sliding time window, and update the risk control prediction network using at least one of incremental learning and federated learning methods to improve the model's adaptive performance in different sites and new environmental scenarios.

Images