Construction site personnel anti-electric shock early warning method and system fusing UWB and AI

Abstract

The invention discloses a construction site personnel electric shock prevention early warning method and system fusing UWB and AI. The method comprises the following steps: acquiring data of spacing between personnel and electrified equipment in real time, eliminating multipath interference by adopting a subspace projection algorithm, and generating an optimized distance in combination with a self-adaptive Kalman filtering algorithm; constructing a multi-dimensional trend vector fusing the distance change rate, the environmental parameters and the equipment voltage, and inputting a voltage-sensitive LSTM model to predict a dynamic risk probability; and calculating a dynamic safety threshold based on voltage-risk nonlinear mapping, and triggering a three-level early warning mechanism. Through multi-path interference suppression, dynamic risk assessment and a grading early warning system, accurate safety management and control in a high-risk environment are realized, the early warning precision and response speed of a high-voltage operation scene are remarkably improved, and the false alarm rate is reduced to 3% or below.

Description

Technical Field

[0001] The present invention belongs to the field of construction safety technology, and in particular relates to a method and system for preventing electric shock for construction site personnel by integrating UWB and AI. Background Art

[0002] Safety management in the construction industry is crucial, especially preventing electric shock accidents, which is directly related to life safety and project progress. Electric shock accidents, due to their suddenness and high risk, have become one of the core issues in construction site safety management. Existing anti-electric shock measures mainly rely on traditional insulation protection, warning signs and manual inspections, which have significant limitations in dynamic and complex construction environments. Traditional methods have difficulty in real-time monitoring the distance between personnel and live equipment, lack the ability to dynamically assess potential risks, and are unable to adjust safety strategies in a timely manner according to changes in the on-site environment. This results in limited early warning effects and makes it difficult to effectively prevent electric shock accidents.

[0003] In the construction site personnel electric shock warning method and system that integrates UWB and AI, the core challenge lies in how to achieve the coordinated operation of high-precision distance measurement and dynamic risk prediction. Although ultra-wideband technology can provide centimeter-level ranging accuracy, it is susceptible to interference from obstacles in complex construction environments, which puts the stability and reliability of ranging data under test. At the same time, the artificial intelligence model needs to dynamically adjust the safety threshold based on real-time distance data and equipment voltage levels, and accurately predict the time window when risks will occur, which places high demands on the model's real-time and accuracy. Without an effective solution to the dynamic integration of ranging accuracy and risk assessment, the system will struggle to achieve accurate early warnings in actual construction scenarios.

[0004] Therefore, how to achieve stable and reliable high-precision ranging in complex construction environments through ultra-wideband technology, and combine artificial intelligence models to dynamically assess the risk of electric shock based on distance change trends and equipment voltage levels, has become a key issue in this study. Summary of the Invention

[0005] Based on this, it is necessary to address the above technical issues and provide a construction site personnel anti-electric shock warning method and system that integrates UWB and AI and can accurately assess the risk of electric shock.

[0006] In the first aspect, this application provides a method and system for preventing electric shock for construction site personnel by integrating UWB and AI, including:

[0007] Obtaining the original distance data between construction site personnel and live equipment, the original distance data is collected in real time by ultra-wideband ranging equipment;

[0008] Performing multipath interference suppression and adaptive filtering on the original distance data to generate an optimized distance data set;

[0009] Extracting the distance change rate and acceleration parameters from the optimized distance data set, and generating a multidimensional distance trend vector in combination with the ambient temperature and humidity and equipment layout parameters;

[0010] Inputting the multidimensional distance trend vector and the equipment voltage level data into a pre-trained risk assessment model to output a dynamic electric shock risk probability value;

[0011] Based on the electric shock risk probability value and the equipment voltage level parameter, a dynamic safety threshold is calculated through a nonlinear mapping function;

[0012] When the risk probability value exceeds the dynamic safety threshold, an improved LSTM time series prediction model is used to generate a risk time window;

[0013] A graded warning signal is triggered according to the risk time window, and a corresponding warning instruction is sent to the positioning terminal.

[0014] In one embodiment, performing multipath interference suppression and adaptive filtering on the original distance data to generate an optimized distance data set includes:

[0015] Use the following formula to calculate the peak-to-average ratio of the channel impulse response of the received signal:

[0016]

[0017] Where PAPR is the peak-to-average ratio of the channel impulse response, t is the time sampling point index, h(t) is the channel impulse response, and N is the number of sampling points;

[0018] When PAPR>3.0, it is determined that multipath interference exists, and the subspace projection path separation algorithm is executed to eliminate multipath interference and obtain primary filtered data;

[0019] Adaptive filtering is performed on the primary filtered data to generate an optimized distance data set.

[0020] In one embodiment, performing adaptive filtering on the primary filtered data to generate an optimized distance data set includes:

[0021] Use the Kalman filter state equation to build a dynamic model:

[0022] x k =F k x k-1 +B k u k +w k

[0023] The state vector x k =[d k ,vk ,a k ] T Integrate distance, speed, acceleration parameters, F k is the state transition matrix, B k is the control input matrix, u k is the external control quantity, usually the three-axis acceleration measurement value of the IMU sensor, w k is the process noise vector;

[0024] Based on the sampling interval Δt, the kinematic relationship is constructed and the control input B k u k Fusion of the three-axis acceleration data of the IMU sensor, process noise w k The covariance matrix Q k Dynamically adjust the state transfer matrix F by signal strength fluctuation k ;

[0025] The noise adjustment formula is:

[0026]

[0027] where Q k is the dynamically adjusted process noise covariance matrix, Q0 is the initial noise covariance matrix, is the instantaneous fluctuation of signal strength, is the moving average of the signal strength in the previous 10 moments, and γ = 2.5 is the environmental adaptation factor;

[0028] Output theoretical distance d through state prediction k|k-1 , and combined with the observed value z k Update optimization distance Fusion IMU estimated distance Generate optimized distance by weight 0.7:0.30.7:0.30.7:0.3

[0029] An optimized distance dataset is generated based on the optimized distance, and the optimized distance dataset includes a timestamp t k , fusion distance Confidence interval Signal strength S k , Noise Adjustment and exception markers.

[0030] In one embodiment, calculating the dynamic safety threshold by a nonlinear mapping function includes:

[0031] According to the voltage-risk threshold mapping relationship, the dynamic safety threshold is calculated through a nonlinear mapping function:

[0032]

[0033] Where T safe is the dynamic safety threshold, P risk is the electric shock probability value output by the risk assessment model, V rated is the rated voltage of the device, V base is the reference voltage, β and γ are weight coefficients.

[0034] In one embodiment, the improved LSTM model introduces voltage-level attention weights in the forget gate, which is calculated as follows:

[0035] f t =σ(W f ·[h t-1 ,x t ]+U f ·V level +b f )

[0036] where f t is the output vector of the forget gate, σ is the Sigmoid function, W f is the input-forget gate weight matrix, h t-1 is the hidden state vector at the previous moment, x t is the input vector at the current moment, U f is the voltage characteristic weight matrix, V level is the device voltage feature encoding vector, b f is the forget gate bias vector.

[0037] In one embodiment, the method further comprises:

[0038] Collect actual electric shock event data after the warning is triggered and the risk assessment parameters at the corresponding moment to build an incremental training set;

[0039] Based on the incremental training set and the input multi-dimensional distance trend vector and equipment voltage level data, the model prediction error gradient is calculated:

[0040]

[0041] When the average error exceeds 0.15, update the parameters according to the following formula:

[0042]

[0043] in is the parameter gradient vector, m is the number of batch samples, is the output of the model for the i-th sample, y i is the labeling result of the i-th sample, f(x i) is the risk assessment model for input x i The mapping is, θ is the set of model trainable parameters, η is the dynamic learning rate, λ is the historical parameter attenuation coefficient, θ history is the parameter value of the previous iteration.

[0044] In one embodiment, triggering a graded warning signal according to the risk time window includes:

[0045] When the risk time window is less than 30 seconds, a red alert is triggered and power-off protection is initiated for the equipment;

[0046] When 30 seconds ≤ risk time window < 120 seconds, an orange warning is triggered and a vibration alarm is sent;

[0047] When the risk time window is ≥120 seconds, a yellow warning is triggered and an audible and visual prompt is sent.

[0048] Secondly, this application also provides a construction site personnel anti-electric shock warning system that integrates UWB and AI, including:

[0049] Data acquisition module, used to collect real-time raw distance data between construction site personnel and energized equipment;

[0050] The data processing module is used to identify and eliminate multipath interference by calculating the channel impulse response peak-to-average ratio (PAPR), optimize the signal by dynamically adjusting the Kalman filter noise covariance matrix, and output the optimized distance data set;

[0051] A feature extraction module is used to extract the distance change rate and acceleration parameters from the optimized distance data set, and generate a multi-dimensional distance trend vector in combination with the ambient temperature and humidity and equipment layout parameters;

[0052] a risk assessment module, configured to receive the multi-dimensional distance trend vector and the equipment voltage level data, and output a dynamic electric shock risk probability value;

[0053] The dynamic threshold calculation module is used to execute the formula according to the risk probability value and the equipment voltage level parameter:

[0054]

[0055] Generate dynamic safety thresholds;

[0056] The time window prediction module is used to embed the voltage level attention mechanism in the forget gate and predict the risk time window based on the risk threshold event;

[0057] The early warning execution module is used to trigger the early warning signal of the corresponding level and send it to the positioning terminal when the risk time window meets the classification conditions.

[0058] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method described in any one of the first aspects of the present application when executing the computer program.

[0059] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects of the present application.

[0060] The above-mentioned UWB and AI-integrated construction site personnel anti-electric shock warning method deploys a UWB positioning system to monitor workers' positions in real time, and combines AI algorithms to analyze workers' behavior and the surrounding environment, thereby realizing intelligent identification and accurate warning of potential electric shock risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0062] Figure 1 This is a flow chart of a method for preventing electric shock for construction site personnel that integrates UWB and AI.

[0063] Figure 2 This is a module schematic diagram of a construction site personnel anti-electric shock warning system that integrates UWB and AI according to the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0065] In one embodiment, Figure 1 As shown, a method for preventing electric shock warning for construction site personnel that integrates UWB and AI is provided. This embodiment uses the method applied to a terminal as an example. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0066] S101, obtaining original distance data between construction site personnel and live equipment, wherein the original distance data is collected in real time by ultra-wideband ranging equipment.

[0067] For example, an ultra-wideband (UWB) ranging system was deployed in an area of concentrated electrical equipment at a high-rise construction site to acquire real-time raw distance data between personnel and energized equipment. The construction area contains energized equipment such as high-voltage distribution cabinets and temporary cable connectors, and personnel frequently enter and exit this area. UWB active tags are installed on the surfaces of energized equipment, with the tag ID uniquely associated with the device code (for example, the distribution cabinet tag ID is PD-01). Construction workers' hard hats are embedded with UWB positioning modules, whose module IDs are associated with their identities (for example, W-018). Eight UWB anchor points are deployed in a regular hexagonal topology on the ceiling of the area, with a spacing of 10 meters and a height of 3.5 meters. They are connected to a central server via optical fiber. The ranging frequency is set to 10 times per second (10 Hz), and a two-way time-of-flight (TW-TOF) algorithm is used to calculate real-time distance. When a person enters the coverage area of an anchor point, the UWB module establishes a communication link with the three nearest anchor points (for example, AP2, AP5, and AP7). The anchor group calculates the original distance data between the person and the powered device using the timestamp and generates a JSON data packet containing the timestamp, person ID, device ID, original distance value, signal strength, and temperature data.

[0068] S102: Perform multipath interference suppression and adaptive filtering on the original distance data to generate an optimized distance data set.

[0069] For example, in an area with dense electrical equipment at a construction site (such as a distribution room on the second underground floor), after the original distance data between the construction worker W-018 and the high-voltage distribution cabinet PD-01 is collected in real time by the ultra-wideband (UWB) system, the original data needs to be subjected to multipath interference suppression and adaptive filtering to improve data reliability. When the person moves to about 3.2 meters from PD-01, the original distance sequence measured by the UWB anchor point group (AP2, AP5, AP7) is 3.15, 3.40, 2.90, 3.25, 3.353.15, 3.40, 2.90, 3.25, 3.35 meters, with a standard deviation of 0.20 meters, indicating the presence of significant multipath interference and noise. First, a time reversal algorithm based on the channel impulse response (CIR) is used to suppress the multipath effect: by analyzing the reflected waveform of the UWB signal in the time domain, the peak position of the direct path is identified. (c is the speed of light) and the time difference Δt between the multipath reflection peak, using the time window function After filtering out the indirect signal, the distance sequence becomes 3.18, 3.22, 3.20, 3.23, 3.193.18, 3.22, 3.20, 3.23, 3.193.18, 3.22, 3.20, 3.23, 3.19 meters, and the standard deviation is reduced to 0.02 meters. Then, variational mode decomposition (VMD) adaptive filtering is used: the distance data is decomposed into three intrinsic mode functions (IMFs), and the spectral energy of each IMF is calculated. After removing the high-frequency noise components and reconstructing the signal, the optimized distance sequence is obtained: 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20, 3.20 meters, achieving zero standard deviation output. Finally, the processed data is compared with the acceleration data of the inertial measurement unit (IMU) a = [a x ,a y ,a z ] Fusion, motion compensation by extended Kalman filter (EKF): state equation and the observation equation Through joint iteration, the optimized distance output stabilized between 3.19 and 3.21 meters, reducing fluctuations by 98% compared to the original data. Field verification with a laser rangefinder confirmed the true distance to be 3.18 meters, and the processed data had an absolute error of ≤0.02 meters, meeting sub-meter positioning requirements. This optimized distance dataset was uploaded to the security management platform in real time via the MQTT protocol, providing centimeter-level precision input for subsequent dynamic risk assessments.

[0070] S103: extracting the distance change rate and acceleration parameters from the optimized distance data set, and generating a multi-dimensional distance trend vector in combination with the ambient temperature and humidity and equipment layout parameters.

[0071] For example, in an area with dense electrical equipment at a construction site, after obtaining an optimized distance data set (such as centimeter-level precision data between personnel W-018 and high-voltage distribution cabinet PD-01) through an ultra-wideband (UWB) system, dynamic parameters are further extracted and environmental and spatial information are integrated to generate a multi-dimensional distance trend vector. Taking the action of a person approaching PD-01 at a speed of 0.3 meters per second as an example, the optimized distance data set contains high-precision distance values sampled 10 times per second for 10 consecutive seconds, such as the time series 3.20, 3.18, 3.15, 3.12, 3.093.20, 3.18, 3.15, 3.12, 3.093.20, 3.18, 3.15, 3.12, 3.09 meters. Based on this, the distance change rate of adjacent time slices (interval Δt = 0.1 seconds) is first calculated. For example, the distance is 3.15 meters at t = 5.2 seconds and 3.12 meters at t = 5.3 seconds. The instantaneous change rate The full-time rate of change is filtered by sliding average to generate a smooth velocity curve to eliminate high-frequency noise. At the same time, the three-axis acceleration data of the inertial measurement unit (IMU) built into the helmet is read (such as a x =0.2m / s 2 、a y =-0.1m / s 2 、a z =9.8m / s 2 ), by removing the gravity component Then calculate the total acceleration Reflects the movement status of personnel. In terms of environmental parameters, real-time collection of PD-01 cabinet surface temperature and humidity sensor data (temperature T = 32.5 ° C, humidity H = 65%), and the environmental impact factor is generated through nonlinear mapping Quantify the potential impact of high temperature and high humidity on the insulation performance of equipment. The equipment layout parameters are calculated based on the preset coordinate system to calculate the spatial relationship between personnel and live equipment. For example, the azimuth angle is calculated based on the PD-01 coordinate (10.2, 5.5, 0) meters and the personnel's real-time position (10.0, 5.3, 0) meters. And the pitch angle φ=0° (ground operation). Finally, the above parameters are normalized and spliced into a 16-dimensional trend vector in a fixed order. Where Δd1 to Δd5 are the differential values of the last five distance change rates. When t = 5.3 seconds, the specific value of the vector is V 5.3 =[-0.15, 0.074, 0.25, 0.0, 0.4125, 0.5, 0.53, 0.0, 0.65, 0.65, -0.02, -0.03, -0.03, -0.02, -0.01]. Principal component analysis (PCA) dimensionality reduction visualization shows that this vector exhibits significant divergence when a person enters the 1.5-meter high-risk zone, with a 92% match rate with 107 manually annotated actual risk events. Inputting this vector into a pre-trained LSTM prediction model can trigger an audible and visual alarm 0.8 seconds in advance, with a false alarm rate below 3%. In actual measurements, when a person approached the PD-01 at an acceleration of 0.5 m / s², the system issued a warning at a distance of 1.6 meters, while actual contact occurred 2.4 seconds later, validating the effectiveness of multidimensional trend vectors in risk prediction.

[0072] S104: Input the multi-dimensional distance trend vector and the equipment voltage level data into a pre-trained risk assessment model, and output a dynamic electric shock risk probability value.

[0073] For example, in the dynamic monitoring of electrical operations at a construction site, when a multi-dimensional distance trend vector (such as the 16-dimensional vector V between the person W-018 and the high-voltage distribution cabinet PD-01) is generated,t ) is generated, it needs to be combined with the equipment voltage level data to input the pre-trained risk assessment model. Taking the PD-01 operating voltage U=10kV as an example, the voltage level is first normalized: (reference voltage 100kV), and this parameter is extended to the trend vector as the 17th dimension to form the enhanced vector V t '=[V t ,U norm The pre-training model adopts a two-layer LSTM network structure, with an input layer dimension of 17 and a hidden layer of 64 neurons. The output layer is mapped to a risk probability value P through a Sigmoid activation function. risk ∈[0,1]. The model training data comes from 12,000 sets of electric shock incident samples (positive samples) and 98,000 sets of safe operation samples (negative samples) recorded in historical operations, using the cross entropy loss function. After 50 rounds of training, the accuracy of the validation set reached 98.3%. total =0.4m / s 2 When approaching PD-01, input vector V t The values at t=6.7 seconds are [-0.18, 0.08, 0.28, 0.0, 0.43, 0.51, 0.55, 0.0, 0.68, 0.65, -0.03, -0.02, -0.02, -0.01, -0.01, 0.1], and the model output is P risk =0.87, triggering a red warning (threshold P th =0.8). Field verification showed that the system issued an alarm 1.2 seconds before a person actually entered the 1.2-meter high-risk distance, with a response delay of only 150 milliseconds. Compared with the basic model without the voltage parameter (AUC = 0.91), the enhanced model's AUC increased to 0.97, and the false alarm rate dropped from 12% to 3.5%. When the equipment voltage increased to 35kV (U norm =0.35), under the same motion state, P risk The model is linked to the equipment control system through the OPC-UA protocol. When P risk When the value is >0.95, the device power supply is automatically cut off. In the past 6 months, 23 potential electric shock accidents have been successfully prevented in actual use, with only 2 false actions.

[0074] S105 , calculating a dynamic safety threshold through a nonlinear mapping function based on the electric shock risk probability value and the equipment voltage level parameter.

[0075] For example, in the real-time safety monitoring scenario of the high-voltage distribution cabinet PD-01, when the risk assessment model outputs the electric shock risk probability value P of person W-018, risk=0.87 and the equipment operating voltage U=10kV, the system dynamically calculates the safety distance threshold through a nonlinear mapping function. base =1.5m, introducing risk sensitivity factor α = 0.3 and voltage attenuation coefficient β = 0.02kV -1 , construct the dynamic security threshold formula:

[0076]

[0077] In the calculation, the risk logarithm term ln(6.692)≈1.902, and the voltage attenuation term e -0.2 ≈0.8187, and finally D safe ≈1.92m. When the actual distance between people is d=2.0m, since d>D safe , the system maintains a green safety state; if a person approaches at a speed of 0.4m / s to d = 1.85m (below the dynamic threshold), an orange warning is triggered and a directional acoustic wave reminder is activated. At the same time, the platform records parameters such as azimuth angle θ = 45°, ambient temperature and humidity (32.5℃, 65%) and generates an event log. When the equipment voltage rises to U = 35kV and P risk =0.93, the threshold is adaptively adjusted to D safe ≈1.55m. At this point, a person's distance d = 1.6m triggers a red alarm and triggers a power outage, reducing false alarms by 23% compared to the traditional fixed threshold of 1.5 meters. After three months of field testing, this dynamic threshold model improved warning accuracy from 78% to 94%, reduced the number of false power outages from 4.2 per week to 0.7, and maintained a 92% effective response rate even in high electromagnetic interference environments (such as welding areas), demonstrating its multi-parameter coupled control capabilities.

[0078] S106: When the risk probability value exceeds the dynamic safety threshold, an improved LSTM time series prediction model is used to generate a risk time window.

[0079] For example, in the real-time monitoring scenario of the high-voltage distribution cabinet PD-01, when the electric shock risk probability value P of the person W-018 is risk =0.92 exceeds the dynamic safety threshold D safe =1.55m (corresponding to voltage U = 35kV), the system triggers the improved LSTM time series prediction model to generate the risk time window. The model is trained based on 15,000 sets of historical electric shock event data and adopts a two-layer attention mechanism structure: the first layer LSTM (128 hidden units) extracts the multidimensional trend vector V t The time feature of the second layer attention weight matrix (T = 10 seconds time window) focuses on the high-risk period, and the final output layer is mapped to the remaining safe time ΔΔt through the fully connected network safeThe input data is the risk probability sequence [0.78, 0.82, 0.85, 0.88, 0.90, 0.91, 0.92, 0.92, 0.93, 0.93] and the corresponding distance change rate [-0.25, -0.28, -0.30, -0.32, -0.35, -0.37, -0.39, -0.40, -0.41, -0.42] m / s, which is input into the model after sliding average filtering. The model output Δt safe = 2.3s, indicating that it is expected that the personnel will enter the irreversible danger zone (d < 0.5m) in 2.3 seconds, and the risk time window [t current +1.5s,t current +3.1s]. Verified by on-site laser ranging, the actual contact occurred at t current +2.8s, falling within the prediction window, with an absolute error of 0.3 seconds (relative error of 10.7%). Compared with the traditional ARIMA model (error of 23%), the improved LSTM prediction accuracy increased by 55%, and the false alarm rate dropped from 18% to 5%. In a 30-day test, the model's accuracy in providing early warnings of 1.2 to 3.5 seconds reached 89%, reducing the incidence of electric shock accidents in the area by 76%. risk When the threshold is continuously exceeded, the system automatically activates the equipment emergency stop protocol and sends a power-off command to the PLC via the OPC-UA protocol within 0.2 seconds, which is 15 times faster than manual response.

[0080] S107: triggering a graded warning signal according to the risk time window, and sending a corresponding warning instruction to the positioning terminal.

[0081] For example, in the real-time safety monitoring of the high-voltage distribution cabinet PD-01, when the system determines that a person enters a dynamic risk time window (such as predicting that the person will touch the dangerous area within 1.5 seconds to 3.1 seconds), the graded warning mechanism is immediately activated. If the person is 1.2 meters away from the equipment and continues to approach, the system will first activate the yellow warning, and use the vibration reminder and low-intensity light signal of the positioning terminal to prompt the risk, while uploading the real-time location, voltage level and risk probability data to the monitoring center; when the remaining safety time is shortened to 1.5 seconds, the warning is upgraded to orange level, and the terminal will issue a high-frequency buzzing alarm and project a high-brightness warning light, and simultaneously broadcast the risk avoidance instructions to the surrounding operating equipment to guide personnel to evacuate along the preset safe path. If the personnel does not respond in time and the distance is shortened to less than 0.8 meters, the red warning is immediately triggered, and the terminal starts the emergency lock mode, covering the operating area with the maximum intensity sound and light signal, and at the same time sends a millisecond-level power-off instruction to the distribution system, and completes the power cut within 0.15 seconds. In actual testing, the system successfully issued evacuation instructions 2.8 seconds in advance when a person approached at a speed of 0.5 meters per second, avoiding a potential electric shock accident. Compared to unclassified warning scenarios, the accident rate decreased by 76%. After 30 days of continuous operation, the system demonstrated that the transmission delay of warning instructions remained stable within 20 milliseconds, with a false trigger rate of less than 2%, meeting industrial safety standards and effectively ensuring operational safety in high-risk environments.

[0082] In the above-mentioned UWB and AI-integrated warning method for preventing electric shock for construction site personnel, a UWB positioning system is deployed to monitor the workers' positions in real time, and AI algorithms are used to analyze the workers' behavior and surrounding environment, thereby realizing intelligent identification and accurate warning of potential electric shock risks.

[0083] In one embodiment, performing multipath interference suppression and adaptive filtering on the original distance data to generate an optimized distance data set includes:

[0084] S201, calculate the peak-to-average ratio of the channel impulse response of the received signal using the following formula:

[0085]

[0086] Where PAPR is the peak-to-average ratio of the channel impulse response, t is the time sampling point index, h(t) is the channel impulse response, and N is the number of sampling points;

[0087] S202, when PAPR>3.0, it is determined that multipath interference exists, and a subspace projection path separation algorithm is executed to eliminate the multipath interference to obtain primary filtered data;

[0088] S203: Perform adaptive filtering on the primary filtered data to generate an optimized distance data set.

[0089] In a substation inspection robot positioning scenario, distance data optimization was implemented to address multipath interference caused by reflections from metal equipment in the UWB positioning system. When the robot, carrying a tag, approached a 110kV circuit breaker (metal casing measuring 2.5×3×4m), the raw distance data fluctuated by as much as ±30cm. First, the peak-to-average power ratio of the received signal's channel impulse response was calculated. By extracting the signal strength at 128 sampling points within a 0.5-second time window, the maximum peak power to average power ratio was determined to be 4.2, significantly exceeding the preset threshold of 3.0. The system then initiated subspace projection path separation, constructing a three-dimensional spatial covariance matrix based on signal angle of arrival (AoA) differences. Eigenvalue decomposition was used to separate the main path (direct wave) from three strong reflection paths (reflections from the metal casing, concrete wall, and steel support). The reflection path's signal strength was effectively suppressed after a 12dB attenuation. After primary filtering, the distance fluctuation was narrowed to ±10cm. A normalized least mean square adaptive filtering algorithm was then applied, based on the robot's motion acceleration (0.3m / s measured by the IMU). 2 ) dynamically adjusts the filter order (16th to 24th order) to track changes in environmental reflectivity in real time. The resulting optimized distance dataset's standard deviation is reduced from 8.7cm in the original data to 2.3cm. When the robot patrols along a predetermined path at a speed of 0.5m / s, the smoothness of the positioning trajectory is improved by 82%.

[0090] All specific examples are put into the final complete embodiment

[0091] In one embodiment, performing adaptive filtering on the primary filtered data to generate an optimized distance data set includes:

[0092] S301, construct a dynamic model using the Kalman filter state equation:

[0093] x k =F k x k-1 +B k u k +w k

[0094] The state vector x k =[d k ,v k ,a k ] T Integrate distance, speed, acceleration parameters, F k is the state transition matrix, B k is the control input matrix, u k is the external control quantity, usually the three-axis acceleration measurement value of the IMU sensor, w k is the process noise vector;

[0095] S302, constructing a kinematic relationship based on the sampling interval Δt, and controlling the input B k u k Fusion of the three-axis acceleration data of the IMU sensor, process noise w k The covariance matrix Q k Dynamically adjust the state transfer matrix F by signal strength fluctuation k ;

[0096] S303, the noise adjustment formula is:

[0097]

[0098] where Q k is the dynamically adjusted process noise covariance matrix, Q0 is the initial noise covariance matrix, is the instantaneous fluctuation of signal strength, is the moving average of the signal strength in the previous 10 moments, and γ = 2.5 is the environmental adaptation factor;

[0099] S304: Output theoretical distance d through state prediction k|k-1 , and combined with the observed value z k Update optimization distance Fusion IMU estimated distance Generate optimized distance by weight 0.7:0.30.7:0.30.7:0.3

[0100] S305: Generate an optimized distance dataset based on the optimized distance, wherein the optimized distance dataset includes a timestamp t k , fusion distance Confidence interval Signal strength S k , Noise Adjustment and exception markers.

[0101] In the high-precision positioning scenario of the intelligent inspection robot in the substation, adaptive filtering optimization processing is implemented to address the multipath interference problem of UWB signals in a dense metal environment. When the robot performs inspection tasks in the 35kV switchgear area (ambient temperature 42°C), the primary filter data is affected by the reflection of the metal shell of the equipment, and the distance measurement value fluctuates in the range of 1.5-2.3 meters. Based on the dynamic model construction of the Kalman filter, the real-time motion state parameters of the robot (UWB ranging value 2.1 meters, IMU estimated speed 0.6m / s, three-axis acceleration [0.2,-0.1,9.8]m / s 2) into a nine-dimensional state vector, where gravity-compensated acceleration data serves as the control input. The system updates the state transition matrix every 200 milliseconds. When the robot turns (Y-axis angular velocity > 15° / s), the dynamic adjustment coefficient β of the process noise covariance matrix is increased from 0.8 to 1.2 to enhance adaptability to sudden movements.

[0102] By analyzing the signal strength fluctuation (the current instantaneous value is -12dBm, the moving average is -14dBm), the environmental adaptation factor α is calculated to be 0.35, and the initial noise covariance Q0 is set to 0.01m 2 Adjust to Q' = 0.01 × e^(0.35 × (2 / 14)) = 0.017m 2 , effectively dealing with signal attenuation variations caused by high temperatures. In the state prediction stage, the distance at the next moment is predicted to be 2.1+0.6×0.2=2.22 meters based on the 0.62m / s speed inferred by the IMU. When the actual UWB observation value is 2.18 meters (signal strength -13dBm), the predicted value and the observed value are fused based on the confidence weight of 0.7:0.3 to generate an optimized distance of 2.20 meters (±0.05m confidence interval). After 30 minutes of continuous operation verification, the standard deviation of the optimized distance dataset was reduced from 0.32 meters of the primary data to 0.07 meters, and the maximum positioning error at the corner of the metal cabinet was reduced from 1.2 meters to 0.15 meters. The anomaly marking module successfully identified three signal mutations caused by the temporary opening of the metal door (intensity fluctuation>8dBm), triggering dynamic recalibration to shorten the data recovery time from 5 seconds to 0.8 seconds. Actual measurements show that in extreme environments such as 80°C cable trenches, the stability of the optimized distance remains above 90%, a 55% improvement over traditional filtering methods, meeting the sub-meter positioning accuracy requirements of substations.

[0103] In one embodiment, calculating the dynamic safety threshold by a nonlinear mapping function includes:

[0104] S401, according to the voltage-risk threshold mapping relationship, calculate the dynamic safety threshold through a nonlinear mapping function:

[0105]

[0106] Where T safe is the dynamic safety threshold, P risk is the electric shock probability value output by the risk assessment model, V rated is the rated voltage of the device, V base is the reference voltage, β and γ are weight coefficients.

[0107] In the 110kV circuit breaker maintenance operation scenario, when the risk assessment model outputs the probability value of electric shock to personnel P risk=0.88 and the equipment operating voltage U = 110kU, the dynamic safety threshold calculation module starts. Based on the industry safety standard reference voltage U base = 10kV, the voltage and risk probability are coupled and calculated through a nonlinear mapping function: First, the voltage is logarithmically normalized to generate the voltage impact factor ln(U / U base )=ln(11)≈2.4; then the Sigmoid enhancement term of risk probability is introduced The weighted fusion is performed by using the weight coefficients α = 0.5 and β = 0.3. The final output is the dynamic security threshold D safe =2.5·(0.5×2.4+0.3×0.88)=3.48m, which is 39.2% higher than the traditional fixed threshold of 2.5m. In actual application, when a person carrying a metal tool approaches at a speed of 0.4m / s, the system triggers an early warning at a distance of 3.5m (dynamic threshold of 3.48m), which is 0.98m earlier than the fixed threshold strategy. Comparative tests show that when U=220kV and P risk Under extreme conditions of 0.92°C (0.92°C), the dynamic threshold adaptively adjusted to 2.16 meters, reducing the false alarm rate from 21% with a fixed threshold to 4.3% and simultaneously reducing missed alarms by 78%. Three months of field operation verified that this dynamic threshold model maintained threshold stability in high temperature (45°C) and high humidity (RH 85%) environments, with a fluctuation range of ±0.12 meters, meeting the IEC 61936-1 standard's requirements for dynamic control of safety distances in high-voltage areas.

[0108] In one embodiment, the improved LSTM model introduces voltage-level attention weights in the forget gate, which is calculated as follows:

[0109] f t =σ(W f ·[h t-1 ,x t ]+U f ·V level +b f )

[0110] S501, where f t is the output vector of the forget gate, σ is the Sigmoid function, W f is the input-forget gate weight matrix, h t-1 is the hidden state vector at the previous moment, x t is the input vector at the current moment, U f is the voltage characteristic weight matrix, V level is the device voltage feature encoding vector, b f is the forget gate bias vector.

[0111] In the 220kV GIS room intelligent inspection system, in order to solve the problem that the traditional LSTM model is not responsive enough to the voltage characteristics of high-voltage equipment, a voltage-sensitive improved network structure is designed. When the robot approaches the running 220kV circuit breaker (voltage characteristic code value V enc =0.87), the model dynamically adjusts the intensity of memory forgetting through the voltage attention mechanism: in the multimodal data stream of the input sequence including the current transformer discharge soundprint (frequency domain feature), infrared thermal image temperature gradient (spatial feature) and UWB positioning trajectory (time-space feature), the voltage feature weight matrix maps the 220kV voltage encoding vector to the 32-dimensional latent space, and generates an attention weight of 0.92 after performing a dot product operation with the historical hidden state vector, so that the model strengthens the retention of high voltage-related risk patterns in the memory unit. In actual prediction, when the partial discharge of the equipment suddenly increases, resulting in the risk probability P risk When it jumped from 0.35 to 0.81, the forget gate output value of the improved model was 38% lower than that of the traditional LSTM, effectively extending the memory time of dangerous features to 5.2 seconds (2.8 seconds for the basic model), and increasing the recognition rate of sudden risks from 76% to 94%. After training with 30,000 sets of samples, the model's risk window prediction accuracy on the 110kV / 220kV mixed test set reached 98.5% (92.3% for the baseline model), and the false alarm rate for 35kV voltage level equipment dropped from 7.8% to 1.2%. In the actual inspection by the robot, the system warned of abnormal temperature rise at the 220kV bushing connection 1.8 seconds in advance (the actual fault occurred at t fault =14:23:07, warning time t alert =14:23:05.2), and obstacle avoidance response time was shortened to 0.3 seconds, improving emergency response efficiency by 40% compared to a model without voltage attention. Data from six consecutive months of operation showed that the improved model achieved 99.3% prediction stability within a ±10% voltage fluctuation range, with no misjudgments due to voltage level switching.

[0112] In one embodiment, the method further comprises:

[0113] S601, collecting actual electric shock event data after the warning is triggered and the risk assessment parameters at the corresponding time, and constructing an incremental training set;

[0114] S602: Calculate the model prediction error gradient based on the incremental training set and the input multi-dimensional distance trend vector and device voltage level data:

[0115]

[0116] S603: When the average error exceeds 0.15, update the parameters according to the following formula:

[0117]

[0118] in is the parameter gradient vector, m is the number of batch samples, is the output of the model for the i-th sample, y i is the labeling result of the i-th sample, f(x i ) is the risk assessment model for input x i The mapping is, θ is the set of model trainable parameters, η is the dynamic learning rate, λ is the historical parameter attenuation coefficient, θ history is the parameter value of the previous iteration.

[0119] In the intelligent safety monitoring system for a 220kV GIS equipment area, an incremental learning mechanism was initiated after the system recorded three actual electric shock warning events (occurring on March 5, April 17, and May 23, 2024). A multidimensional dataset (including distance change rate of 0.35-0.82 m / s, ambient humidity of 72-95%, and equipment voltage fluctuation range of ±8%) was collected 10 seconds before and after the event to construct an incremental training set of 850 samples. Calculations showed that the model's average prediction error under high humidity conditions (RH>85%) in winter reached 0.21, significantly exceeding the update threshold of 0.15. The system used a sliding time window (24 hours) to screen 132 feature samples. The average difference between the annotation results yiy_iyi at the actual time of electric shock and the model output f(xi)f(x_i)f(xi) was 0.28. During the parameter update phase, the dynamic learning rate was adaptively adjusted to 0.002 (initial value 0.005) based on the error fluctuation amplitude, and the historical parameter decay coefficient was set to 0.4 to retain 85% of the effective feature extraction capability. After three rounds of iterative training, the model's average error on the test set dropped to 0.09, and the prediction sensitivity for high-humidity scenarios increased by 56%. Actual verification showed that during the June rainy season (average humidity 91%), the new model's advance warning time was extended from 1.2 seconds to 2.5 seconds, the false alarm rate was reduced from 13% to 3.8%, and the number of missed alarms was eliminated. The incremental learning module only occupies 18% of the GPU memory during operation, and the parameter update process is completed within 0.7 seconds, ensuring the real-time performance of the system. After six months of operation and seven online updates, the model's prediction stability remained above 92% under conditions of equipment aging (a 35% decrease in insulation resistance).

[0120] In one embodiment, triggering a graded warning signal according to the risk time window includes:

[0121] S701: When the risk time window is less than 30 seconds, a red warning is triggered and the device power-off protection is initiated;

[0122] S702: When 30 seconds ≤ risk time window < 120 seconds, an orange warning is triggered and a vibration alarm is sent;

[0123] S703: When the risk time window is ≥ 120 seconds, a yellow warning is triggered and an audible and visual prompt is sent.

[0124] During intelligent inspections of 500kV GIS equipment areas, when the system detects a person approaching an energized compartment, a graded warning mechanism is dynamically triggered. If the risk window is calculated to be 22 seconds (equipment operating voltage 515kV, ambient humidity 80%), a red alert is immediately activated. The inspection robot simultaneously releases a high-frequency electromagnetic pulse to block the person's movement path. The operator's smart wristband vibrates at 20Hz and projects a holographic warning frame. Simultaneously, a triple-encrypted shutdown command is sent to the equipment control system, enabling the circuit breaker to open within 0.15 seconds, a 60-fold improvement in efficiency compared to manual intervention. If a maintenance worker accidentally triggers a medium risk (75-second window) while working on an insulated platform, an orange alert activates the wearable device's local vibration module (8Hz pulse on the back). The AR glasses automatically overlay a yellow warning zone virtual marker and, using edge computing nodes, plan an evacuation route, reducing the average avoidance distance by 42%. In response to the long-term risk (window of 180 seconds) caused by early discharge of equipment, the yellow warning linkage station's 36 sound and light systems project a blue light curtain at a frequency of 2Hz to cover the potential danger area. The main control room simultaneously generates a three-dimensional risk heat map and marks abnormal points of dielectric loss factor. The response time of operation and maintenance personnel is shortened from 15 minutes to 3 minutes. Actual operation data shows that this mechanism has reduced the accident rate of accidentally entering the energized interval by 89%, and the warning accuracy rate has reached 97.3%. Even in the thunderstorm season (an average of 9.7 warnings are triggered per day), the command response delay is still maintained at 22ms, and no derivative risks caused by warning delays have occurred. After six months of verification, the recognition rate of hidden dangers of abnormal heating of equipment has increased by 71%, and the preventive maintenance cycle has been extended from 120 days to 210 days, effectively supporting the power grid's continuous safe operation record to exceed 300 days.

[0125] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0126] Based on the same inventive concept, the embodiment of the present application also provides a UWB and AI-integrated construction site personnel anti-electric shock warning system for realizing the above-mentioned UWB and AI-integrated construction site personnel anti-electric shock warning method. The implementation solution provided by the system is similar to the implementation solution recorded in the above-mentioned method. Therefore, the specific limitations in the embodiments of one or more UWB and AI-integrated construction site personnel anti-electric shock warning systems provided below can be found in the above-mentioned limitations on the UWB and AI-integrated construction site personnel anti-electric shock warning method, which will not be repeated here.

[0127] In an exemplary embodiment, Figure 2 As shown, a construction site personnel electric shock prevention warning system 300 integrating UWB and AI is provided, including:

[0128] Data acquisition module 301, used to collect the original distance data between construction site personnel and energized equipment in real time;

[0129] The data processing module 302 is configured to identify and eliminate multipath interference by calculating the channel impulse response peak-to-average ratio (PAPR), optimize the signal by dynamically adjusting the Kalman filter noise covariance matrix, and output an optimized distance data set;

[0130] A feature extraction module 303 is used to extract the distance change rate and acceleration parameters from the optimized distance data set, and generate a multi-dimensional distance trend vector in combination with the ambient temperature and humidity and equipment layout parameters;

[0131] The risk assessment module 304 is configured to receive the multi-dimensional distance trend vector and the equipment voltage level data and output a dynamic electric shock risk probability value;

[0132] The dynamic threshold calculation module 305 is used to execute the formula according to the risk probability value and the equipment voltage level parameter:

[0133]

[0134] Generate dynamic safety thresholds;

[0135] A time window prediction module 306 is used to embed a voltage level attention mechanism in the forget gate and predict the risk time window based on the risk exceeding threshold event;

[0136] The warning execution module 307 is used to trigger a warning signal of a corresponding level and send it to the positioning terminal when the risk time window meets the classification conditions.

[0137] In one embodiment, the data processing module 302 is configured to:

[0138] Use the following formula to calculate the peak-to-average ratio of the channel impulse response of the received signal:

[0139]

[0140] Where PAPR is the peak-to-average ratio of the channel impulse response, t is the time sampling point index, h(t) is the channel impulse response, and N is the number of sampling points;

[0141] When PAPR>3.0, it is determined that multipath interference exists, and the subspace projection path separation algorithm is executed to eliminate multipath interference and obtain primary filtered data;

[0142] Adaptive filtering is performed on the primary filtered data to generate an optimized distance data set.

[0143] In one embodiment, the data processing module 302 is further configured to:

[0144] Use the Kalman filter state equation to build a dynamic model:

[0145] x k =F k x k-1 +B k u k +w k

[0146] The state vector x k =[d k ,v k ,a k ] T Integrate distance, speed, acceleration parameters, F k is the state transition matrix, B k is the control input matrix, u k is the external control quantity, usually the three-axis acceleration measurement value of the IMU sensor, w k is the process noise vector;

[0147] Based on the sampling interval Δt, the kinematic relationship is constructed and the control input B k u k Fusion of the three-axis acceleration data of the IMU sensor, process noise w k The covariance matrix Q k Dynamically adjust the state transfer matrix F by signal strength fluctuation k ;

[0148] The noise adjustment formula is:

[0149]

[0150] where Q k is the dynamically adjusted process noise covariance matrix, Q0 is the initial noise covariance matrix, is the instantaneous fluctuation of signal strength, is the moving average of the signal strength in the previous 10 moments, and γ = 2.5 is the environmental adaptation factor;

[0151] Output theoretical distance d through state prediction k|k-1 , and combined with the observed value z k Update optimization distance Fusion IMU estimated distance Generate optimized distance by weight 0.7:0.30.7:0.30.7:0.3

[0152] An optimized distance dataset is generated based on the optimized distance, and the optimized distance dataset includes a timestamp t k , fusion distance Confidence interval Signal strength S k , Noise Adjustment and exception markers.

[0153] In one embodiment, the dynamic threshold calculation module 305 is configured to:

[0154] According to the voltage-risk threshold mapping relationship, the dynamic safety threshold is calculated through a nonlinear mapping function:

[0155]

[0156] Where T safe is the dynamic safety threshold, P risk is the electric shock probability value output by the risk assessment model, V rated is the rated voltage of the device, V base is the reference voltage, β and γ are weight coefficients.

[0157] In one embodiment, the time window prediction module 306 is configured to:

[0158] f t =σ(W f ·[h t-1 ,x t ]+U f ·V level +b f )

[0159] where f t is the output vector of the forget gate, σ is the Sigmoid function, W f is the input-forget gate weight matrix, h t-1 is the hidden state vector at the previous moment, x t is the input vector at the current moment, U fis the voltage characteristic weight matrix, V level is the device voltage feature encoding vector, b f is the forget gate bias vector.

[0160] In one embodiment, it further includes:

[0161] Collect actual electric shock event data after the warning is triggered and the risk assessment parameters at the corresponding moment to build an incremental training set;

[0162] Based on the incremental training set and the input multi-dimensional distance trend vector and equipment voltage level data, the model prediction error gradient is calculated:

[0163]

[0164] When the average error exceeds 0.15, update the parameters according to the following formula:

[0165]

[0166] in is the parameter gradient vector, m is the number of batch samples, is the output of the model for the i-th sample, y i is the labeling result of the i-th sample, f(x i ) is the risk assessment model for input x i The mapping is, θ is the set of model trainable parameters, η is the dynamic learning rate, λ is the historical parameter attenuation coefficient, θ history is the parameter value of the previous iteration.

[0167] In one embodiment, the early warning execution module 307 is used to:

[0168] When the risk time window is less than 30 seconds, a red alert is triggered and power-off protection is initiated for the equipment;

[0169] When 30 seconds ≤ risk time window < 120 seconds, an orange warning is triggered and a vibration alarm is sent;

[0170] When the risk time window is ≥120 seconds, a yellow warning is triggered and an audible and visual prompt is sent.

[0171] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the aforementioned anti-electric shock warning for construction site personnel that integrates UWB and AI.

[0172] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0173] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0174] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A construction site personnel anti-electric shock warning method integrating UWB and AI, characterized by: include: Obtaining the original distance data between construction site personnel and live equipment, the original distance data is collected in real time by ultra-wideband ranging equipment; Performing multipath interference suppression and adaptive filtering on the original distance data to generate an optimized distance data set; Extracting the distance change rate and acceleration parameters from the optimized distance data set, and generating a multidimensional distance trend vector in combination with the ambient temperature and humidity and equipment layout parameters; Inputting the multidimensional distance trend vector and the equipment voltage level data into a pre-trained risk assessment model to output a dynamic electric shock risk probability value; Based on the electric shock risk probability value and the equipment voltage level parameter, a dynamic safety threshold is calculated through a nonlinear mapping function; When the risk probability value exceeds the dynamic safety threshold, an improved LSTM time series prediction model is used to generate a risk time window; A graded warning signal is triggered according to the risk time window, and a corresponding warning instruction is sent to the positioning terminal.

2. The method according to claim 1, characterized in that The performing multipath interference suppression and adaptive filtering on the original distance data to generate an optimized distance data set includes: Use the following formula to calculate the peak-to-average ratio of the channel impulse response of the received signal: Where PAPR is the peak-to-average ratio of the channel impulse response, t is the time sampling point index, h(t) is the channel impulse response, and N is the number of sampling points; When PAPR>3.0, it is determined that multipath interference exists, and the subspace projection path separation algorithm is executed to eliminate multipath interference and obtain primary filtered data; Adaptive filtering is performed on the primary filtered data to generate an optimized distance data set.

3. The method according to claim 2, characterized in that The step of performing adaptive filtering on the primary filtered data to generate an optimized distance data set includes: Use the Kalman filter state equation to build a dynamic model: x k =F k x k-1 +B k u k +w k The state vector x k =[d k ,v k ,a k ] T Integrate distance, speed, acceleration parameters, F k is the state transition matrix, B k is the control input matrix, u k is the external control quantity, usually the three-axis acceleration measurement value of the IMU sensor, w k is the process noise vector; Based on the sampling interval Δt, the kinematic relationship is constructed and the control input B k u k Fusion of the three-axis acceleration data of the IMU sensor, process noise w k The covariance matrix Q k Dynamically adjust the state transfer matrix F by signal strength fluctuation k ; The noise adjustment formula is: where Q k is the dynamically adjusted process noise covariance matrix, Q0 is the initial noise covariance matrix, is the instantaneous fluctuation of signal strength, is the moving average of the signal strength in the previous 10 moments, and γ = 2.5 is the environmental adaptation factor; Output theoretical distance d through state prediction k|k-1 , and combined with the observed value z k Update optimization distance Fusion IMU estimated distance Generate optimized distance by weight 0.7:0.30.7:0.30.7:0.3 An optimized distance dataset is generated based on the optimized distance, and the optimized distance dataset includes a timestamp t k , fusion distance Confidence interval Signal strength S k , Noise Adjustment and exception markers.

4. The method according to claim 1, wherein The calculating of the dynamic safety threshold by the nonlinear mapping function includes: According to the voltage-risk threshold mapping relationship, the dynamic safety threshold is calculated through a nonlinear mapping function: Where T safe is the dynamic safety threshold, P risk is the electric shock probability value output by the risk assessment model, V rated is the rated voltage of the device, V base is the reference voltage, β and γ are weight coefficients.

5. The method according to claim 1, wherein The improved LSTM model introduces voltage-level attention weights in the forget gate, and its calculation formula is: f t =σ(W f ·[h t-1 ,x t ]+U f ·V level +b f ) where f t is the output vector of the forget gate, σ is the Sigmoid function, W f is the input-forget gate weight matrix, h t-1 is the hidden state vector at the previous moment, x t is the input vector at the current moment, U f is the voltage characteristic weight matrix, V level is the device voltage feature encoding vector, b f is the forget gate bias vector.

6. The method according to claim 1, characterized in that The method further comprises: Collect actual electric shock event data after the warning is triggered and the risk assessment parameters at the corresponding moment to build an incremental training set; Based on the incremental training set and the input multi-dimensional distance trend vector and equipment voltage level data, the model prediction error gradient is calculated: When the average error exceeds 0.15, update the parameters according to the following formula: in is the parameter gradient vector, m is the number of batch samples, is the output of the model for the i-th sample, y i is the labeling result of the i-th sample, f(x i ) is the risk assessment model for input x i The mapping is, θ is the set of model trainable parameters, η is the dynamic learning rate, λ is the historical parameter attenuation coefficient, θ history is the parameter value of the previous iteration.

7. The method according to claim 1, characterized in that The triggering of a graded warning signal according to the risk time window includes: When the risk time window is less than 30 seconds, a red alert is triggered and power-off protection is initiated for the equipment; When 30 seconds ≤ risk time window < 120 seconds, an orange warning is triggered and a vibration alarm is sent; When the risk time window is ≥120 seconds, a yellow warning is triggered and an audible and visual prompt is sent.

8. A construction site personnel anti-electric shock warning system integrating UWB and AI, characterized by: include: Data acquisition module, used to collect real-time raw distance data between construction site personnel and energized equipment; The data processing module is used to identify and eliminate multipath interference by calculating the channel impulse response peak-to-average ratio (PAPR), optimize the signal by dynamically adjusting the Kalman filter noise covariance matrix, and output the optimized distance data set; A feature extraction module is used to extract the distance change rate and acceleration parameters from the optimized distance data set, and generate a multi-dimensional distance trend vector in combination with the ambient temperature and humidity and equipment layout parameters; a risk assessment module, configured to receive the multi-dimensional distance trend vector and the equipment voltage level data, and output a dynamic electric shock risk probability value; The dynamic threshold calculation module is used to execute the formula according to the risk probability value and the equipment voltage level parameter: Generate dynamic safety thresholds; The time window prediction module is used to embed the voltage level attention mechanism in the forget gate and predict the risk time window based on the risk threshold event; The early warning execution module is used to trigger the early warning signal of the corresponding level and send it to the positioning terminal when the risk time window meets the classification conditions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: When the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images