An electric power field operation AI safety early warning method, system and head-mounted device

Abstract

The application belongs to the technical field of power operation safety early warning, and provides a power field operation AI safety early warning method, system and head-mounted device, which synchronously collects multi-dimensional data of a power operation field, at least including electric field intensity data, operation personnel physiological state data, operation environment video data and motion posture data; the obtained operation environment video data is analyzed in real time to identify power exclusive dangerous targets; the electric field intensity data and the operation personnel physiological state data are coupled to judge and decouple features to obtain near-electricity decoupling features and heart rate decoupling features; the preliminary risk is determined in combination with the power exclusive dangerous target identification result, the motion posture data and the decoupled features; the final model input features are obtained in combination with the normalized initial risk score and the feature weighted fusion after the fusion of various features; the real-time risk level is output based on the final model input features; and the corresponding early warning trigger strategy is generated according to the real-time risk level and the set early warning trigger mechanism. The application changes from "after-the-fact tracing" to "intervention in the middle", and is suitable for power multi-scene risk differences.

Description

Technical Field

[0001] This invention belongs to the field of power operation safety early warning technology, and in particular relates to an AI-based safety early warning method, system and head-mounted device for power field operations. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of the power industry, the working scenarios faced by front-line power service personnel are becoming increasingly complex, creating an urgent need for safety management, efficient collaboration, and intelligent assistance during operations. Especially in on-site operations such as substation inspections, line maintenance, and equipment installation, workers are often exposed to multiple potential risks such as high voltage, strong electromagnetic fields, and falls from heights. Traditional safety management methods that rely on visual observation and experience-based judgment are no longer sufficient to meet the requirements of modern power safety production.

[0004] To address these challenges, head-mounted devices have begun to be applied in the field of power field operations. Existing solutions mainly fall into two categories: one is the traditional safety helmet with only physical protection functions, lacking any intelligent sensing and early warning capabilities; the other is an integrated smart safety helmet, which integrates modules such as cameras and walkie-talkies, enabling basic video recording and voice communication. However, these integrated devices usually suffer from problems such as limited functionality, inconvenient interaction, and limited computing power. Their early warnings mostly rely on backend manual monitoring and cannot achieve real-time, local intelligent intervention.

[0005] To improve functionality and wearing flexibility, split-type designs have emerged. For example, the computing control unit is separated from the head-mounted display unit. However, current split-type designs have limited AI analysis capabilities or rely on the cloud. General models have low accuracy in identifying power-specific hazards such as insulator cracks and loose grounding wires. Furthermore, the "end-to-cloud" collaborative mode suffers from output transmission delays, and in environments with weak networks or strong electromagnetic interference, such as substations, real-time performance and reliability cannot be guaranteed. Local computing power is insufficient to support real-time inference of complex models, and uploading multimodal sensor data, including biometrics, to the cloud poses a privacy risk and does not comply with the stringent data security compliance requirements of the power industry. The split-type hangers often use simple hook designs, which are difficult to adapt to safety helmets of different brands and curvatures, resulting in poor contact between sensors (such as proximity sensors) and the head or helmet surface, leading to large data collection errors and affecting the accuracy of warnings. In summary, existing head-mounted devices suffer from poor functional and scenario adaptability, potential risks in computing architecture and data security, and poor device compatibility. Summary of the Invention

[0006] In order to solve at least one of the technical problems existing in the background art, the present invention provides an AI safety early warning method, system and head-mounted device for power field operations, which upgrades from "post-event traceability" to "in-event intervention" and is adapted to the risk differences of multiple power scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides an AI-based safety early warning method for power field operations, applied to a head-mounted device, comprising the following steps: Simultaneously collect multi-dimensional data from the power operation site, including at least electric field strength data, physiological status data of the workers, video data of the working environment, and motion posture data; Real-time analysis of acquired video data of the working environment to identify power-specific hazardous targets; By performing coupling degree judgment and feature decoupling on electric field strength data and physiological state data of workers, near-electric decoupling features and heart rate decoupling features are obtained; A preliminary risk assessment is made by combining the identification results of power-specific hazardous targets, motion posture data, and decoupled features. The final model input features are obtained by combining the normalized initial risk score and the feature weighted fusion after the fusion of various features. The real-time risk level is output based on the final model input features. Based on the real-time risk level and the set early warning triggering mechanism, a corresponding early warning triggering strategy is generated.

[0008] Furthermore, the real-time analysis of the acquired operational environment video data to identify power-specific hazardous targets includes: The preprocessed video data of the working environment is input into the input layer for downsampling, and then processed by CLAHE illumination normalization to obtain standardized input data; Standardized input data is fed into the backbone network and processed through a newly added power frequency interference suppression channel, the Swish activation function, and customized batch normalization parameters to obtain multi-scale feature maps. Multi-scale feature maps are input into the neck network, and small target features are enhanced by pre-defined multiple types of power-specific anchor boxes to obtain fused features. The fused features are input into the head network to activate the core power category, and the power-specific hazardous target identification results are output using a dynamic confidence threshold.

[0009] Furthermore, let the near-field electrical intensity sequence be X and the heart rate sequence be Y. Interpolate Y to a sequence Y' with the same frequency as X, and then use the mutual information entropy between X and Y'. Perform coupling degree determination, when If the coupling value exceeds the preset coupling threshold, it is determined to be a strong coupling. The Cross-Attention mechanism is then activated to decouple the features of X and Y, resulting in decoupled near-electrical decoupling features and heart rate decoupling features, respectively.

[0010] Furthermore, the near-electrical decoupling characteristics and heart rate decoupling characteristics after decoupling are expressed as follows: Near-electrical decoupling characteristics: , The heart rate decoupling characteristics are: , in, This refers to the near-electrical decoupling characteristics after decoupling. The decoupled heart rate characteristics are shown below. It is a near-electrical intensity sequence. The heart rate sequence Y is interpolated to a sequence with the same frequency as X. and These are the attention weights for near-electrical features and the attention weights for heart rate features, respectively.

[0011] Furthermore, the risk assessment based on motion posture data and decoupled near-electrical characteristics and heart rate characteristics yields risk assessment results, including: The decoupled features are standardized to obtain the corresponding standardized features; Based on the electrical work safety regulations, and combined with the establishment of a rule layer, a preliminary risk assessment is conducted on three types of standardized characteristics, and a preliminary risk score is output. ; Determine the weight coefficients of the extracted features and normalize the preliminary risk score. The features are then fused with various other features to obtain a weighted fusion of the fused features, which is used to obtain the final model input features. : Input the final model with features The final risk level is output by combining SVM model reasoning.

[0012] Furthermore, the preliminary risk assessment, which combines the results of power-specific hazardous target identification, motion posture data, and decoupled features, includes: When the condition is met and and When the target recognition module does not detect a preset dangerous target, it determines... ; When the condition 0 < is satisfied ≤0.2、 When =0 and P_norm=0, determine =1; When condition 0.2 < ≤0.6 and 0< If the value is ≤0.25, or if the target recognition module detects that the safety helmet is not worn correctly, then the determination is made. =2; When the target recognition module detects "not wearing a safety helmet", or meets the condition 0.2 < ≤0.6, or meet the condition 0.25< When ≤0.5, determine =3; When condition 0.6 < ≤0.8 and 0.5< ≤1, or when the target recognition module detects that the insulating gloves are not being worn, a determination is made. =4; When the condition is met >0.8 and >0.5, or meet the conditions When =1, determine =5; where, if there are cases that satisfy multiple rule conditions, then the multiple rule conditions corresponding to... The maximum value among the values ​​is used as the final risk score. ; in, For near-electric distance standardized parameters, For heart rate standardized parameters, Parameters for predicting falls from heights.

[0013] Furthermore, the method also includes locally encrypting and storing the collected metadata and processing results according to a hierarchical encryption strategy, specifically including: For heart rate and body temperature biometric data, the national cryptographic algorithm SM7 is used for encryption; For operational video data, encryption is differentiated according to the real-time risk level: video frames during risk warning periods are encrypted using the national cryptographic standard SM7, while video frames during normal periods are encrypted using AES-128; the specified encrypted video data is decrypted and uploaded only upon receiving an authorization instruction containing an electronic signature. Metadata such as risk level, timestamp, and device number are automatically uploaded in plaintext with digital signatures when the risk level exceeds a preset threshold.

[0014] A second aspect of the present invention provides an AI-based safety early warning system for power field operations, comprising: The data acquisition module is used to synchronously collect multi-dimensional data from the power operation site, including at least electric field strength data, physiological state data of the workers, video data of the work environment, and motion posture data. The target recognition module is used to analyze the acquired video data of the working environment in real time and identify power-specific hazardous targets. The feature decoupling module is used to determine the coupling degree of electric field strength data and worker physiological state data and decouple them to obtain near-electric decoupling features and heart rate decoupling features. The risk assessment module is used to make a preliminary risk judgment by combining the identification results of power-specific dangerous targets, motion posture data and decoupled features. It combines the normalized initial risk score and the weighted fusion of various features to obtain the final model input features, and outputs the real-time risk level based on the final model input features. The early warning module is used to generate corresponding early warning triggering strategies based on the real-time risk level and the set early warning triggering mechanism.

[0015] A third aspect of the present invention provides a head-mounted device, including an AI safety early warning system for power field operations as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention creatively proposes an AI-based safety early warning method for power field operations, and develops an AI-based safety early warning system and head-mounted device for power field operations. It collects and integrates multi-dimensional core data from power operations through multi-source data fusion technology, and combines customized AI algorithms to achieve accurate identification of power-specific hazardous targets, decoupling of proximity to electricity and heart rate features, and multi-dimensional risk assessment. By adopting a dynamic early warning triggering strategy, it realizes integrated collaboration across the entire process of power field operation data collection, target identification, risk assessment, and real-time early warning. This significantly improves the real-time performance and accuracy of power field operation safety early warnings, ensuring the safety and stability of power field operations. It adapts to the different risk levels in various power scenarios, solving the problems of poor scenario adaptability and assessment bias caused by feature coupling in current power field operation safety early warning systems.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of an AI-based safety early warning method for power field operations provided by an embodiment of the present invention; Figure 2 This is a schematic diagram of the adjustable curvature hook assembly provided in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] Example 1 like Figure 1 As shown, this embodiment provides an AI-based safety early warning method for power field operations, including the following steps: Step 1: Synchronously collect multi-dimensional data from the power operation site, including at least electric field strength data, physiological status data of the workers, video data of the working environment, and motion posture data; In this embodiment, multi-dimensional data from the power operation site are synchronously collected based on the multi-modal sensor module installed on the head-mounted device, including at least proximity sensors, biosensors, image-stabilized cameras, pressure sensors, and accelerometers; Specifically, the proximity sensor: collects electric field strength signals at 100Hz and converts them using a 12-bit ADC (accuracy ±0.01kV / m). Biosensors: collect heart rate at 1Hz and body temperature at 0.5Hz, and transmit data via I²C interface; Image-stabilized camera: captures video stream at 30fps, resolution 1280×720, H.264 encoding; Pressure sensor: 1Hz to collect contact pressure; accelerometer: 100Hz to collect motion data. All data is synchronized via hardware timestamps (1ms precision).

[0024] Step 2: Preprocess the acquired multidimensional data; In this embodiment, the preprocessing of the acquired multidimensional data includes: For near-electric signals, moving average filtering, low-pass filtering, and power frequency interference band-stop filtering are used for processing; For video stream data, frame alignment + image stabilization compensation + grayscale conversion + histogram equalization; For sensor data, linear interpolation of missing values ​​(sensor anomaly alarm triggered when more than 5 consecutive missing sampling points are detected), 3σ removal of outliers, and Cross-Attention decoupling of strongly coupled data are employed.

[0025] Step 3: Based on a customized target detection AI model for the power scenario, perform real-time analysis of the acquired video data of the working environment to identify power-specific hazardous targets; In this embodiment, the target detection AI model customized for the power scenario is an Elec-YOLO model based on the improved YOLO architecture. The Elec-YOLO model is compressed using knowledge distillation technology to achieve real-time inference on the NPU of the computing power processing unit (130).

[0026] Specifically, the steps include the following: Step 301: Input the preprocessed work environment video data into the input layer for downsampling, such as... The resolution was adjusted to 640×640 (downsampled from the original 1280×720), the image format was converted to YUV420, and the image was normalized using CLAHE illumination normalization to obtain standardized input data. Step 302: Input the standardized input data into the backbone network, and process it through the newly added power frequency interference suppression channel, Swish activation function and customized batch normalization parameters to obtain multi-scale feature maps; Among them, the newly added power frequency interference suppression channel adopts a 3×3 convolution kernel, which is 1.2 times the number of the original model. The Swish activation function replaces the original LeakyReLU. The customized batch normalization parameters are based on statistics of 10,000 power scene maps, with mean values ​​of R=123.68, G=116.28, and B=103.94. Step 303: Input the multi-scale feature map into the neck network (PANet), and enhance the fusion of small target features through preset multi-class power-specific anchor boxes to obtain the fused features; In this embodiment, three types of power-specific anchor frames are preset: near-electric arc: 15×15, 30×30, 45×45; insulator crack: 20×50, 30×70, 40×90; exposed copper wiring: 25×25, 50×50, 75×75.

[0027] Step 304: Input the fused features into the head network, activate only the 6 core power categories (near-electric hazard, insulator crack, exposed copper wiring, not wearing a safety helmet, falling objects from height, and not wearing insulating gloves), and use dynamic confidence thresholds (threshold 0.65 when near-electric strength > 1kV / m; threshold 0.8 when near-electric strength ≤ 0.5kV / m) to output the power-specific hazard target identification results.

[0028] Step 4: Determine the coupling degree and decouple the electric field strength data and the physiological state data of the workers; In this embodiment, let the near-field electrical intensity sequence be X, and the heart rate sequence be Y. Y is interpolated to a sequence Y' with the same frequency as X; the mutual information entropy between X and Y' is calculated. : , , , , in, for X Information entropy Let Y' be the information entropy. for and Joint information entropy, The first in the near-electrical intensity sequence i One data point, For the heart rate sequence i One data point, For the heart rate sequence j One data point, Statistical probability for power scenario samples; The criteria for judgment are: If I(X,Y')>the preset coupling threshold, it is determined to be a strong coupling. The Cross-Attention mechanism is activated to decouple the features of X and Y, and the decoupled near-electrical decoupling features and heart rate decoupling features are obtained respectively. If I(X,Y')≤ preset coupling threshold, proceed directly to the original PCA dimensionality reduction process of the document.

[0029] In this embodiment, the coupling threshold can be set to 0.6; The Cross-Attention mechanism is activated to decouple the features of X and Y, resulting in decoupled near-electrical decoupling features and heart rate decoupling features, specifically including: Near-electrical decoupling characteristics: , The heart rate decoupling characteristics are: , in, , These are the attention weights for near-electrical features and heart rate features, respectively, and their dimensions are 15×15 in this implementation. By performing coupling degree judgment and feature decoupling on electric field strength data and worker physiological state data as described above, the detection error of proximity detection and the error of heart rate variability coefficient are eliminated; it has been verified that the proximity detection error is reduced from ±0.5 meters to ±0.2 meters, and the heart rate variability coefficient (HRV) error is reduced by 40%.

[0030] Step 5: Based on motion posture data and decoupled proximity characteristics and heart rate characteristics, perform risk assessment to obtain risk assessment results. Combine the power-specific hazardous target identification results and risk assessment results to output the real-time risk level. Specifically, the steps include the following: Step 501: Standardize the decoupled features to obtain the corresponding standardized features; Since the three types of features have different physical dimensions (near-electric feature is distance / electric field strength, heart rate feature is bpm, and attitude feature is acceleration / angle), they are first standardized to unify the data range to the [0,1] interval to avoid the evaluation results being dominated by a single feature with an excessively large dimension. Near-Electricity Decoupling Characteristic Standardization: Let the original value of the near-electricity decoupling characteristic be... (Corresponding to the near-electric distance D, unit: meters, calculated from electric field strength data), the standardized formula is: , The conversion relationship between proximity distance and electric field strength is based on a preset calibration curve: D=k / E, where k is the scenario calibration coefficient, obtained by fitting 100 sets of measured data from power operation sites. ≥0.98; Heart rate decoupling feature standardization: Let the original value of the heart rate decoupling feature be... (Unit: bpm), the standardized formula is: , Behavioral posture feature standardization: Behavioral posture features are three-dimensional accelerations collected by accelerometers ( ) and attitude angles (pitch angle θ, roll angle) The attitude risk coefficient is first calculated using the following formula, obtained by fusing the yaw angle (ψ) and the yaw angle (ψ). : , Where g = 9.8 m / s² is the acceleration due to gravity, and the attitude angle is in degrees (divided by 100 to balance the magnitudes of acceleration and angle) before standardization: , Step 502: Based on the electrical work safety regulations, establish a rule layer to conduct preliminary risk assessment of the three types of standardized features and output preliminary risk scores. ; In this embodiment, the specific processing rules include: like and and Meanwhile, the target recognition module did not detect dangerous targets such as "not wearing a safety helmet" or "not wearing insulated gloves," and thus determined... (Corresponding to Level 0); If 0 < ≤0.2 (close to the power source at a distance of 3-5 meters) and =0 and =0, judge =1 (corresponding to Level 1); If 0.2 < ≤0.6 (close to the electric field at a distance of 2-3 meters) and 0 < If the heart rate is ≤0.25 (80-100 bpm), or the target recognition module detects "improper helmet wearing", then the following judgment is made. =2 (corresponding to Level 2); If the target recognition module detects "helmet not being worn", or 0.2 < ≤0.6, or 0.25< ≤0.5 (heart rate 100-120 bpm) is considered =3 (corresponding to Level 3); If 0.6 < ≤0.8 (close to the power source at a distance of 1-2 meters) and 0.5 < ≤1 (heart rate 120-140 bpm), or the target recognition module detects "insulating gloves not being worn", then a judgment is made. =4 (corresponding to Level 4); like >0.8 (near-electric distance <1 meter) and >0.5, or =1 (High-altitude fall prediction), judgment =5 (corresponding to Level 5); If a scenario with overlapping rules occurs (e.g., "close to electric shock at 2 meters + heart rate 110 bpm + helmet not worn"), the maximum score from the rule layer will be taken as the result. .

[0031] Step 503: Determine the weight coefficients of the extracted features and normalize the preliminary risk score. The features are then fused with various other features to obtain a weighted fusion of the fused features, which is used to obtain the final model input features. : The formula for calculating the fused feature vector F is as follows: , in, Near-electrical normalized feature weights For heart rate standardized feature weights, The pose-normalized feature weights are wE=0.5, wHR=0.3, and wP=0.2. The various features are fused to obtain the fused features, which are then compared with the preliminary risk score. After normalization Considering the different weights of the three types of characteristics on the risk of power operations (near-power characteristics are the core risk source with the highest weight; heart rate characteristics are an auxiliary indicator of human condition; posture characteristics are a supplementary indicator of behavioral risk), the weight coefficients are determined using the Analytic Hierarchy Process (AHP): Final model input features : , 0.6 and 0.4 are the feature fusion weights for the model input, determined by cross-validation of 500 labeled samples to ensure a balance between rule priors and data features. These values ​​can be set according to actual needs. Step 504: Input the final model features. The final risk level is output by combining the improved SVM model inference. In this embodiment, the improved SVM model uses the radial basis function (RBF) to adapt to the nonlinear correlation of the three types of features in the power scenario. The kernel function formula is as follows: , where γ=0.8 (obtained by grid search optimization); Introducing a feature attention layer adds attention weight adjustment before the SVM input layer. The weight of the near-electrical related components is dynamically increased (when... When the value is greater than 0.5, the attention weighting coefficient increases to 1.2, strengthening the impact of core risk characteristics; The penalty coefficient was optimized by setting the penalty coefficient C=10 to balance the classification accuracy of the model for positive examples (dangerous scenarios) and negative examples (safe scenarios), thereby reducing the false negative rate in high-risk scenarios.

[0032] During training, the improved SVM model refines the risk level labels output by the SVM model, comparing them with the initial scores from the rule layer. Perform a consistency check. If the two are consistent, output the risk level directly; otherwise, if there are discrepancies (e.g., at the rule level),... If the model output is 4, then the model output will be used as the standard (because the model incorporates the nonlinear correlation features of real-time data), and the difference samples will be stored locally for subsequent model iteration and optimization.

[0033] Step 6: Generate corresponding early warning triggering strategies based on the real-time risk level and the set early warning triggering mechanism; In this embodiment, the early warning triggering mechanism refers to a "layered response + dynamic adaptation" mechanism built based on risk level (Level 0-Level 5), personnel proficiency, and the urgency of the scenario. Its core includes four key elements: triggering conditions, early warning intensity, interaction method, and emergency backup. This ensures that early warnings in different risk scenarios are accurate without interfering with operations. The specific strategy is as follows: Level 0: Normal (distance to power > 3 meters, heart rate 60-100 bpm, helmet worn correctly, no specific electrical hazards). Triggering condition: Triggered by no-risk feature Warning strategy: Do not activate audible, visual, or vibration alarms; only record the work status in the local log (e.g., "Current work environment is safe, physiological state is normal"), which can be viewed in real time on the backend management platform without the need for operator intervention.

[0034] Level 1: Low risk (close to electricity 3-5 meters, heart rate 60-100 bpm, normal working posture, no electrical hazards) Triggering condition: Low-risk proximity characteristics detected, with no other hazardous associations. Warning strategy: Only activate voice reminder (TTS synthesis, volume 60dB, to avoid interrupting the operation), no lights or vibration; The unified message for beginners, experienced users, and experts is: "You are currently in a low-risk near-electricity zone. Please maintain a safe distance." This message is broadcast once and lasts for 1 second.

[0035] Level 2: Low to medium risk (close to electric shock at a distance of 2-3 meters, heart rate of 80-100 bpm, or improper helmet wearing) Triggering conditions: The proximity to the electrical system is close to the safety threshold, or there is a slightly non-compliant target. Warning strategy: Activate weak sound and light alerts + repeated voice prompts, without vibration; Lighting: Green light flashes slowly (1 time / second), no red light involved; Voice prompt (volume 70dB): Beginner: "Low to medium risk warning: close to electric current / safety helmet not worn correctly. Please adjust your work position and wear protective equipment correctly" (repeated twice, 3 seconds apart). Proficient / Expert: "Low to medium risk, please maintain a safe distance / wear your helmet properly" (repeated 1 time).

[0036] Level 3: Medium to high risk (not wearing a helmet / close to an electric shock at a distance of 2-3 meters / heart rate of 100-120 bpm) Triggering conditions: Existence of clear violations or medium-risk sources Warning strategy: Activate standard sound and light alarm + targeted voice prompts, and add slight vibration (intensity 30%) for beginners. Audible and visual alarm: Red and blue lights flash alternately (frequency 2 times / second), alarm volume 85dB, continuously triggered until the risk is cleared; Voice prompts (TTS synthesis, dynamically generated based on risk type): Situation where no safety helmet is worn: "Medium-high risk! No safety helmet detected. Please stop work immediately and put on a safety helmet." Near-electric shock scenario: "Medium to high risk! The distance to the electric shock is too close. Please immediately move back to a safe area at least 3 meters away." High heart rate scenario: "Medium to high risk! Your heart rate is abnormally high. Please stop working and adjust your state before continuing." Layered response to avoid a "one-size-fits-all" approach: Level 0-2 targets "potential risks" and "minor violations," using light warnings (logs only, low-volume voice, dim lighting) to avoid interfering with normal operations; Level 3-5 targets "clear risks" and "fatal threats," gradually increasing the warning intensity (overlay of sound, light, and vibration, emergency response linkage) to ensure that risks are detected in a timely manner. In electrical work, the urgency of the risks is different between "being 3 meters away from electricity" (Level 1) and "not wearing a safety helmet" (Level 3). The classification of levels allows early warnings to accurately match the severity of the risks. All status levels are recorded locally, and the backend can analyze historical data to "the pattern of low-risk scenarios turning into medium- and high-risk scenarios" (such as the process of the proximity distance to electricity gradually decreasing from 4 meters to 2 meters), providing data support for the subsequent optimization of the early warning mechanism.

[0037] User adaptation: Beginners will experience slight vibration (once every 2 seconds, for 3 times), while experienced / expert users will only receive sound and light + voice.

[0038] Level 4: High risk (close to electric shock at a distance of 1-2 meters + heart rate of 120-140 bpm, or not wearing insulated gloves) Triggering conditions: High-risk sources combined with abnormal human body conditions, or lack of critical protective measures; Early warning strategy: Activate high-intensity audible, visual and vibration alarms + mandatory operation prompts, regardless of personnel proficiency (all enhanced early warnings); Sound and light: Red and blue lights flash rapidly (3 times / second), alarm volume is 90dB (5dB higher than medium and high risk), continuous alarm; Vibration: High-intensity vibration (70% intensity), continuously triggered until the risk is eliminated; Voice prompts (repeated in a loop, 2-second intervals): Scenario of near-electric shock + abnormal heart rate: "High risk! You have entered a dangerous near-electric shock zone and your heart rate is high. Immediately turn off the tool power and quickly retreat to at least 3 meters away!"; Situation where insulated gloves are not worn: "High risk! No insulated gloves detected. Do not touch electrical equipment. Please put on gloves immediately before working!"

[0039] Level 5: Extremely high risk (close electrical distance < 1 meter + heart rate > 120 bpm / prediction of fall from height) Triggering conditions: Deadly risk source or urgent threat to personal safety Early warning strategy: Activate the highest level of early warning + automatic emergency linkage, and simultaneously trigger the SOS assistance function; Sound and light: Red and blue lights flash (5 times / second), alarm volume 95dB (maximum volume), without interruption; Vibration: Extremely strong vibration (100% intensity), continuous vibration; Voice prompt (rapidly repeating): "Extremely high risk! Evacuate immediately! Evacuate immediately!"; Emergency Response: Automatically records the current location (positioning accuracy ±1 meter), encrypts and records on-site video, and simultaneously pushes an "extremely high risk alarm" to the backend management platform. The platform automatically pops up a reminder to the administrator, and new users can additionally trigger a "one-click help" voice prompt ("Activate SOS? Press and hold the SOS button to quickly call for support").

[0040] Personnel proficiency matching (new): Based on locally stored job records (early warning response accuracy, hazard handling time), it is divided into 3 levels: Beginners (less than 3 months of experience): Audible and visual alarms + vibration + detailed operating instructions (such as "Danger of electric shock, please back up immediately. Operating steps: 1. Turn off the tool power; 2. Keep away from high-voltage equipment"). Proficient (3-24 months): Audible and visual alarm + brief prompt (e.g., "Danger near electric shock, back away"); Expert (>24 months): Audible and visual alarm only; SOS button trigger: Press and hold for 2 seconds to activate SOS mode, simultaneously start recording (encrypted storage), location upload, and call the management platform; short press triggers red and blue warning lights.

[0041] Step 7: The collected raw data and processing results are encrypted and stored locally according to a hierarchical encryption strategy, and the data that has been anonymized or de-identified is uploaded to the backend management platform when the upload conditions are met; In this embodiment, a "power-level encryption + authorized upload" mechanism is adopted. The original data is encrypted and stored locally (with a maximum retention period of 7 days), and only metadata is uploaded as needed, which complies with the requirements of the "Power Data Security Management Measures".

[0042] For heart rate and body temperature biometric data, the national cryptographic algorithm SM7 is used for encryption; For operational video data, encryption is differentiated according to the real-time risk level: video frames during risk warning periods are encrypted using the national cryptographic standard SM7, while video frames during normal periods are encrypted using AES-128; the specified encrypted video data is decrypted and uploaded only upon receiving an authorization instruction containing an electronic signature. Metadata such as risk level, timestamp, and device number are automatically uploaded in plaintext with digital signatures when the risk level exceeds a preset threshold.

[0043] Example 2 This embodiment provides an AI-based safety early warning system for power field operations, including: The data acquisition module is used to synchronously collect multi-dimensional data from the power operation site, including at least electric field strength data, physiological state data of the workers, video data of the work environment, and motion posture data. The target recognition module is used to analyze the acquired video data of the working environment in real time and identify power-specific hazardous targets. The feature decoupling module is used to determine the coupling degree of electric field strength data and worker physiological state data and decouple them to obtain near-electric decoupling features and heart rate decoupling features. The risk assessment module is used to perform risk assessment based on motion posture data and decoupled near-electricity characteristics and heart rate characteristics, obtain risk assessment results, and output real-time risk level by combining the power-specific hazardous target identification results and risk assessment results. The early warning module is used to generate corresponding early warning triggering strategies based on the real-time risk level and the set early warning triggering mechanism.

[0044] It should be noted that the specific implementation of the AI ​​safety early warning system for power field operations in this embodiment of the invention is similar to the specific implementation of the AI ​​safety early warning method for power field operations in this embodiment of the invention. Please refer to the description in the method section for details. In order to reduce redundancy, it will not be repeated here.

[0045] Example 3 like Figure 2 As shown, this embodiment provides a head-mounted device, which is a split structure, including a hanging recorder 1, a first bracket 2, a second bracket 3, and a helmet body 4. The first bracket 2 is symmetrically fastened to both sides of the hanging recorder 1 to form a hanging assembly. The second bracket 3 is used in conjunction with the hanging assembly. One end of the second bracket 3 is fastened to both sides of the helmet body 4 with different curvatures, and the other end is fastened and fixed to the first bracket 2 on the hanging assembly.

[0046] Furthermore, the under-mounted recorder is U-shaped with adjustable curvature, and the second bracket is compatible with helmets of different curvatures, so the entire assembly can be detachably adapted and installed on helmets of different curvatures.

[0047] The head-mounted device includes: a multimodal sensor module for synchronously acquiring multidimensional data from the work site, including at least a proximity sensor for detecting electric field strength, a biosensor for acquiring personnel physiological states, a stabilized camera for acquiring environmental video, and an accelerometer for acquiring motion posture; and a computing power processing unit with a built-in heterogeneous computing architecture centered on an NPU and locally deployed with a customized target detection AI model for the power scenario. The computing power processing unit is connected to the multimodal sensor module via an internal bus to receive the multidimensional data and perform the following operations: Run a target detection AI model customized for power scenarios to analyze the video stream captured by the image-stabilized camera in real time and identify power-specific dangerous targets; An attention-based feature decoupling algorithm is invoked to determine the coupling degree and decouple the electric field intensity data collected by the proximity sensor from the physiological data collected by the biosensor, so as to eliminate the interference of physiological state on proximity detection. Based on the decoupled features, target recognition results, and accelerometer data, a real-time risk level is output through a risk assessment algorithm. A safety early warning module, connected to the computing power processing unit, is used to trigger corresponding sound, light, and vibration alarms based on the real-time risk level. The voice interaction unit integrates a locally deployed speech recognition model and a power wake-up-free command library, which is used to recognize and respond to voice commands in specific power operation scenarios without the need for a wake word, and provide feedback through speech synthesis. The data storage and encryption module adopts a hierarchical encryption strategy to locally encrypt and store the raw data and processing results collected by the multimodal sensor module, and only uploads authorized or de-identified data to the background management platform when the upload conditions are met.

[0048] It should be noted that the specific implementation of the head-mounted device in this embodiment of the invention is similar to the specific implementation of the AI ​​safety early warning method for power field operations in this embodiment of the invention. Please refer to the description in the method section for details. To reduce redundancy, it will not be repeated here.

[0049] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the AI-based safety early warning method for power field operations as described above.

[0050] Example 4 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the AI ​​safety early warning method for power field operations as described above.

[0051] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0052] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0053] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0054] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0055] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0056] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for AI-based safety early warning in power field operations, applied to a head-mounted device, characterized in that, include: Simultaneously collect multi-dimensional data from the power operation site, including at least electric field strength data, physiological status data of the workers, video data of the working environment, and motion posture data; Real-time analysis of acquired video data of the working environment to identify power-specific hazardous targets; By performing coupling degree judgment and feature decoupling on electric field strength data and physiological state data of workers, near-electric decoupling features and heart rate decoupling features are obtained; A preliminary risk assessment is made by combining the identification results of power-specific hazardous targets, motion posture data, and decoupled features. The final model input features are obtained by combining the normalized initial risk score and the feature weighted fusion after the fusion of various features. The real-time risk level is output based on the final model input features. Based on the real-time risk level and the set early warning triggering mechanism, a corresponding early warning triggering strategy is generated.

2. The AI-based safety early warning method for power field operations as described in claim 1, characterized in that, The real-time analysis of acquired video data of the working environment to identify power-specific hazardous targets includes: The preprocessed video data of the working environment is input into the input layer for downsampling, and then processed by CLAHE illumination normalization to obtain standardized input data; Standardized input data is fed into the backbone network and processed through a newly added power frequency interference suppression channel, the Swish activation function, and customized batch normalization parameters to obtain multi-scale feature maps. Multi-scale feature maps are input into the neck network, and small target features are enhanced by pre-defined multiple types of power-specific anchor boxes to obtain fused features. The fused features are input into the head network to activate the core power category, and the power-specific hazardous target identification results are output using a dynamic confidence threshold.

3. The AI-based safety early warning method for power field operations as described in claim 1, characterized in that, a Let X be the near-field electrical intensity sequence and Y be the heart rate sequence. Interpolate Y to a sequence Y' with the same frequency as X. Then, based on the mutual information entropy of X and Y'... Perform coupling degree determination, when If the coupling value exceeds the preset coupling threshold, it is determined to be a strong coupling. The Cross-Attention mechanism is then activated to decouple the features of X and Y, resulting in decoupled near-electrical decoupling features and heart rate decoupling features, respectively.

4. The AI-based safety early warning method for power field operations as described in claim 3, characterized in that, The near-electrical decoupling characteristics and heart rate decoupling characteristics after decoupling are expressed as follows: Near-electrical decoupling characteristics: , The heart rate decoupling characteristics are: , in, This refers to the near-electrical decoupling characteristics after decoupling. The decoupled heart rate characteristics are shown below. It is a near-electrical intensity sequence. The heart rate sequence Y is interpolated to a sequence with the same frequency as X. and These are the attention weights for near-electrical features and the attention weights for heart rate features, respectively.

5. The AI-based safety early warning method for power field operations as described in claim 1, characterized in that, The risk assessment based on motion posture data and decoupled near-electrical characteristics and heart rate characteristics yields the following results: The decoupled features are standardized to obtain the corresponding standardized features; Based on the electrical work safety regulations, and combined with the establishment of a rule layer, a preliminary risk assessment is conducted on three types of standardized characteristics, and a preliminary risk score is output. ; Determine the weight coefficients of the extracted features and normalize the preliminary risk score. The features are then fused with various other features to obtain a weighted fusion of the fused features, which is used to obtain the final model input features. : Input the final model with features The final risk level is output by combining SVM model reasoning.

6. The AI-based safety early warning method for power field operations as described in claim 5, characterized in that, The preliminary risk assessment, which combines the results of power-specific hazardous target identification, motion posture data, and decoupled features, includes: When the condition is met and and When the target recognition module does not detect a preset dangerous target, it determines... ; When the condition 0 < is satisfied ≤0.2、 When =0 and P_norm=0, determine =1; When condition 0.2 < ≤0.6 and 0< If the value is ≤0.25, or if the target recognition module detects that the safety helmet is not worn correctly, then the determination is made. =2; When the target recognition module detects "not wearing a safety helmet", or meets the condition 0.2 < ≤0.6, or meet the condition 0.25< When ≤0.5, determine =3; When condition 0.6 < ≤0.8 and 0.5< ≤1, or when the target recognition module detects that the insulating gloves are not being worn, a determination is made. =4; When the condition is met >0.8 and >0.5, or meet the conditions When =1, determine =5; where, if there are cases that satisfy multiple rule conditions, then the multiple rule conditions corresponding to... The maximum value among the values ​​is used as the final risk score. ; in, For near-electric distance standardized parameters, For heart rate standardized parameters, Parameters for predicting falls from heights.

7. The AI-based safety early warning method for power field operations as described in claim 1, characterized in that, The method also includes locally encrypting and storing the collected metadata and processing results according to a hierarchical encryption strategy, specifically including: For heart rate and body temperature biometric data, the national cryptographic algorithm SM7 is used for encryption; For operational video data, encryption is differentiated according to the real-time risk level: video frames during risk warning periods are encrypted using the national cryptographic standard SM7, while video frames during normal periods are encrypted using AES-128; the specified encrypted video data is decrypted and uploaded only upon receiving an authorization instruction containing an electronic signature. Metadata such as risk level, timestamp, and device number are automatically uploaded in plaintext with digital signatures when the risk level exceeds a preset threshold.

8. A power field operation AI safety early warning system, characterized in that, include: The data acquisition module is used to synchronously collect multi-dimensional data from the power operation site, including at least electric field strength data, physiological state data of the workers, video data of the work environment, and motion posture data. The target recognition module is used to analyze the acquired video data of the working environment in real time and identify power-specific hazardous targets. The feature decoupling module is used to determine the coupling degree of electric field strength data and worker physiological state data and decouple them to obtain near-electric decoupling features and heart rate decoupling features. The risk assessment module is used to make a preliminary risk judgment by combining the identification results of power-specific dangerous targets, motion posture data and decoupled features. It combines the normalized initial risk score and the weighted fusion of various features to obtain the final model input features, and outputs the real-time risk level based on the final model input features. The early warning module is used to generate corresponding early warning triggering strategies based on the real-time risk level and the set early warning triggering mechanism.

9. A head-mounted device, characterized in that, This includes the AI-based safety early warning system for power field operations as described in claim 8.

10. A head-mounted device as described in claim 9, characterized in that, The headgear has a split structure, including an adjustable curvature hook assembly and a safety helmet body. The adjustable curvature hook assembly can be detachably adapted and installed on the power safety helmet.

Images