Safety belt unhooking prevention monitoring method and system in tower climbing process of electric power operation personnel

Abstract

The invention discloses a safety belt unhooking prevention monitoring method and system in the tower climbing process of an electric power operator, and relates to the technical field of electric power safety monitoring, and the method comprises the following steps: S1, unhooking monitoring sensor deployment and reference data calibration; s2, collecting and preprocessing unhooking monitoring data; s3, performing multi-dimensional intelligent analysis on the unhooking behavior; s4, unhooking risk grading early warning response is carried out; according to the monitoring method and system for preventing the safety belt from unhooking in the tower climbing process of the electric power operator, through an LSTM dynamic risk prediction model, the limitation of traditional'post alarm 'is broken through, the future unhooking probability can be predicted in advance, and precious risk avoiding time is won for the operator; and in combination with an improved SVM classification model, three unhooking states of'normal, suspected and confirmed 'are accurately distinguished, and operation interruption caused by misjudgment or falling accidents caused by missed judgment are avoided.

Description

Technical Field

[0001] This application relates to the field of power safety monitoring technology, and more specifically, to a method and system for monitoring the safety belt disengagement of power workers during tower climbing. Background Technology

[0002] Against the backdrop of rapid development in the power industry and continuous advancements in smart grid technology, tower climbing operations, as a core component of power grid equipment maintenance and repair, are directly related to the safety of personnel and the stable operation of the power system. The industry has therefore imposed stringent standardization requirements on tower climbing safety. However, tower climbing operations often face complex environments such as high altitudes, strong winds and sandstorms, and large temperature differences, and existing safety harness systems have significant technical shortcomings.

[0003] Insufficient data processing accuracy: Relying solely on a single sensor to monitor key connection points makes it difficult to accurately capture data related to decoupling behavior. Furthermore, existing filtering algorithms mostly use fixed parameters and cannot adaptively eliminate dynamic interference such as wind and sand impacts and tower vibrations, resulting in a large error rate in the identification of decoupling events.

[0004] Lack of risk prediction capabilities: Existing technologies can only achieve passive alarm after decoupling, and cannot predict risk trends based on historical data and real-time characteristics. The early warning response time is too long, missing the best time to avoid risks.

[0005] Therefore, in response to the aforementioned technical issues, a monitoring method and system for preventing the safety belt from becoming detached during the process of power workers climbing towers is proposed. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, this application provides a method and system for monitoring the safety belt decoupling of power workers during tower climbing, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this application provides the following technical solution: a method for monitoring the disengagement of safety belts during the process of power workers climbing towers, the method comprising the following steps:

[0008] S1. Decoupling Monitoring Sensor Deployment and Baseline Data Calibration: MEMS sensors are selected and installed at key risk points of the seat belt. Baseline data calibration is completed through zero-point calibration and multi-scenario simulation.

[0009] S2. Decoupling monitoring data acquisition and preprocessing: Real-time acquisition of sensor physical signals, followed by outlier removal, data completion and adaptive filtering to obtain clean and valid data;

[0010] S3. Multi-dimensional intelligent analysis of decoupling behavior: Extract decoupling sensitive features through feature engineering, fuse multi-sensor data based on improved DS evidence theory, use improved SVM to classify decoupling status, combine LSTM model to predict dynamic risk trends, and finally complete risk level quantification through fuzzy comprehensive evaluation.

[0011] S4. Risk Level Early Warning Response for Decoupling: Based on the risk level, trigger local alarms, remote push notifications and emergency response measures to provide a level early warning system, while also enabling visualized monitoring of the work scenario based on a digital twin model.

[0012] By adopting the above technical solutions and through the integration and innovation of multiple technologies, core issues such as sensor adaptability, data accuracy, risk prediction, and visual monitoring can be solved.

[0013] Furthermore, the MEMS sensor in S1 includes a triaxial accelerometer, a tension sensor, and a displacement sensor; the key risk points of the safety belt include the connection between the main hook and the safety rope, the connection between the auxiliary hook and the tower body fixing point, and the tension monitoring point in the middle section of the safety rope. Specifically, at the connection between the main hook and the safety rope, an accelerometer and a tension sensor are installed to monitor the sudden acceleration and tension drop when the hook falls off.

[0014] At the connection point between the auxiliary hook and the tower body: install displacement and acceleration sensors to monitor hook displacement and vibration.

[0015] Tension monitoring point in the middle section of the safety rope: Install a tension sensor to help verify the overall stress state of the safety rope;

[0016] The sensor sensitivity threshold is adjusted based on the decoupling mechanical characteristic curve; the sensitivity adjustment formula is as follows: ,in Sensitivity; Initial sensitivity; , These represent the maximum and minimum tensions during the unhooking process; To improve the accuracy of target tension detection.

[0017] By adopting the above technical solution, a bonding method is used during installation. High-strength epoxy resin is used to tightly bond the sensor to the connection point, with an installation gap of ≤0.1mm, eliminating signal acquisition deviation caused by gaps. At the same time, the sensor cable adopts an anti-pull design with a tensile strength of ≥500N to prevent cable breakage during operation.

[0018] Furthermore, the specific strategy for completing the reference data calibration through zero-point calibration in S1 is as follows: place the sensor in a static environment, collect 100 sets of static data, calculate the mean as the zero-point reference, and eliminate the initial data error.

[0019] The specific strategy for completing the benchmark data calibration through multi-scenario simulation is as follows: record the sensor response data in each scenario and establish a benchmark threshold library for three types of states: "normal state, suspected decoupling, and confirmed decoupling".

[0020] By adopting the above technical solution, based on the stress analysis of the safety belt and the distribution of decoupling risk points, sensors are precisely deployed at three key locations: the main hook-safety rope, the auxiliary hook-tower fixing point, and the middle section of the safety rope. The sensors are fixed in a close fit (installation gap ≤ 0.1mm) and use anti-pull cables (tensile strength ≥ 500N). This avoids interference with normal operation by sensor installation and ensures complete collection of key decoupling signals, covering all risk points in the tower climbing operation process.

[0021] Furthermore, the adaptive filtering process in S2 employs an improved Kalman filter algorithm, and the filter update formula is as follows:

[0022] State prediction: ;

[0023] Covariance prediction: ;

[0024] Kalman gain: ,in Variance is updated in real time via a sliding window. , The length of the window;

[0025] Status Update: ;

[0026] Covariance update: ;

[0027] in, for Predicted state value at any given time; for State value after filtering at any given time; for State value after filtering at any given time; This is the state transition matrix; For control matrix; To control the quantity; It is the covariance matrix; for Time-prediction error covariance matrix; for Time error covariance matrix; For process noise covariance; The observation matrix; for Time-based observations; To observe the noise covariance; It is the identity matrix; for Time-based Kalman gain; Find the inverse of a matrix; This is the matrix transpose.

[0028] By adopting the above technical solution, the sensor collects data in real time. The collection parameters are set as follows: sampling frequency: 100Hz for acceleration sensor, 50Hz for tension sensor, and 20Hz for displacement sensor.

[0029] Data collected includes: acceleration (x, y, z axes), tension, displacement, and temperature and humidity data of the sensor itself.

[0030] Data transmission: Data is transmitted to the edge computing gateway via Bluetooth 5.0 protocol.

[0031] Preliminary screening of the collected raw data:

[0032] Outlier removal: The 3σ criterion is used; Data completion: If there is missing data (i.e., the missing rate is ≤5%), linear interpolation is used to complete the data; Validity verification: Data that conforms to the fluctuation range of the seat belt under normal working conditions is retained, such as the normal range of acceleration ±1g and the normal range of displacement ±3mm, to ensure that the data used for analysis is complete and reliable.

[0033] To address environmental interference during tower climbing operations, such as wind and sand impact (high-frequency interference, 100Hz~500Hz) and tower vibration (low-frequency interference, 1Hz~10Hz), an improved Kalman filter algorithm is employed: State vector definition. ,in , , It is a triaxial acceleration. For tension, For displacement;

[0034] State transition matrix Observation matrix Process noise covariance ,in It is a diagonal matrix;

[0035] Window length After filtering, the signal-to-noise ratio of the data is improved, effectively removing the influence of environmental interference on the real signal.

[0036] Furthermore, the specific strategy for improving the DS evidence theory fusion of multi-sensor data in S3 is as follows:

[0037] Introducing sensor reliability weights ,and ,in For the first Historical misjudgment rate of each sensor Number of sensors;

[0038] Conflict evidence processing employs a weighted average adjusted BPA value, the adjustment formula being: , To identify a subset of states in the framework, For the first Each sensor pair The original BPA value, For the first Each sensor pair The corrected BPA value; the fusion formula is .

[0039] By adopting the above technical solutions and using an improved Kalman filtering algorithm (with real-time updates of observation noise covariance), the signal-to-noise ratio of the data is improved, effectively filtering out dynamic interferences such as wind and sand impacts and tower vibrations. Combined with the improved DS evidence theory that introduces sensor reliability weights, the judgment bias caused by interference from a single sensor is solved. After the fusion of multi-source data, the global confidence error is effectively reduced, providing a high-purity data foundation for subsequent state classification.

[0040] Furthermore, the improved SVM decoupling state classification in S3 specifically involves: using the Grey Wolf Optimization Algorithm (GWO) to optimize the penalty coefficient of the SVM. With kernel function parameters Optimize the objective function for: ,in For real labels, For predicting labels, For the number of samples; SVM classification decision function for ,in For the number of support vectors, For Lagrange multipliers, It is a radial basis kernel function, and , For bias terms, It is a symbolic function.

[0041] Furthermore, the dynamic risk prediction input based on the LSTM model in S3 is a sequence of sensitive features over the past five time windows. The "historical input data" of the LSTM model, consisting of "decoupling-sensitive features" from nearly five time windows, forms the basis for predicting future risks. Given a set of multidimensional feature vectors (e.g., 12-dimensional, including decoupling-sensitive features such as peak acceleration of the main / secondary hooks, variance of safety rope tension, and proportion of high-frequency energy), the output is the risk probability for the next two time windows. ,in For the next time window ( The probability of a definite decoupling occurring; For the next time window ( Given the probability of a definite decoupling, the state update strategy of the LSTM model is as follows:

[0042] Forgotten Gate: ;

[0043] Input Gate: ;

[0044] Cell state: ;

[0045] Output gate: ;

[0046] in, , , These are the forget gate, input gate, and output gate, respectively. , These represent the current cell state and the current hidden state, respectively. , These represent the cell state at the previous moment and the hidden state at the previous moment, respectively. , , , This is the weight matrix; , , , For bias terms; It is the sigmoid function; This is element-wise multiplication; This represents the current state of the candidate cells; It is the hyperbolic tangent function.

[0047] The LSTM model structure consists of 2 LSTM layers (64 neurons per layer) + 1 fully connected layer (32 neurons) + 1 output layer (2 neurons), with activation functions of tanh (LSTM layer), ReLU (fully connected layer), and sigmoid (output layer).

[0048] Model training: Adam optimizer was used, with a learning rate of 0.001, a batch size of 32, 100 iterations, and cross-entropy loss function. ,in For batch sample size, For true probability, To predict probabilities;

[0049] Predictive applications: If predicting This will trigger an early warning, giving workers ≥0.3 seconds to avoid danger.

[0050] Furthermore, the specific strategy for quantifying risk levels through fuzzy comprehensive evaluation in S3 is as follows:

[0051] Specifically: Select the displacement deviation rate Duration of acceleration exceeding the limit Tension decrease range Integrating Trust LSTM predicted risk value To evaluate the indicators, the indicator weights are determined using the analytic hierarchy process (AHP), and the consistency test formula is as follows: , , To determine the largest eigenvalue of a matrix, For random consistency index; fuzzy evaluation matrix Comprehensive evaluation results , For the indicator weight vector, according to The final risk level is determined by the fuzzy level corresponding to the maximum value in the middle.

[0052] By adopting the above technical solutions, through LSTM dynamic prediction and multi-model cross-validation (such as SVM probability threshold determination), the invalid early warning rate is reduced, frequent work interruptions caused by false alarms are avoided, tower climbing efficiency is improved, and the maintenance and downtime costs of power equipment are indirectly reduced.

[0053] Furthermore, a monitoring system for preventing safety belt decoupling during the tower climbing process of power workers according to the method described above, the system comprising:

[0054] The sensing layer consists of MEMS sensors, a positioning module, and temperature and humidity sensors, and is used to collect physical signals and environmental information.

[0055] Data layer: includes edge computing gateway and cloud database. The edge computing gateway performs data preprocessing, and the cloud database stores raw data, benchmark data and analysis results.

[0056] Analysis Layer: Deploys an improved DS evidence theory fusion module, an improved SVM classification module, an LSTM prediction module, and a fuzzy comprehensive evaluation module to complete multi-dimensional intelligent analysis. Among them, the improved DS evidence theory fusion module is used to perform improved DS evidence theory fusion of multi-sensor data; the improved SVM classification module achieves decoupled state classification through improved SVM; the LSTM prediction module combines an LSTM model to predict dynamic risk trends; and the fuzzy comprehensive evaluation module completes risk level quantification through fuzzy comprehensive evaluation.

[0057] Application layer: includes local alarm terminals, monitoring center platform, and digital twin visualization module, to realize hierarchical early warning and visual monitoring.

[0058] The digital twin visualization module uses Unity3D to build a 3D model of the tower climbing operation scene, synchronizes the perception layer data in real time, and updates the model at a frequency of 10Hz. It supports personnel location tracking, visualization of safety belt connection status, and highlighting of risk areas. The emergency linkage interface can be connected to the power operation management system. When a level 3 warning is triggered, it automatically pushes the personnel location, decoupling risk point, and emergency response plan to the management system, and initiates the rescue dispatch process. The interface communication protocol uses MQTT.

[0059] The above technical solution uses Unity3D to construct a 1:1 three-dimensional model of the tower climbing operation, and synchronizes the sensing layer data at a frequency of 10Hz to achieve real-time tracking of the operator's position, color marking of the safety belt connection status (green - normal, yellow - suspected, red - unhooked), and highlighting of risk areas. Monitoring personnel can intuitively grasp the operation status, and click on the sensor markers to view real-time data curves and historical playback, solving the problems of "data abstraction and scene fragmentation" in traditional monitoring, and is especially suitable for remote supervision and operation guidance.

[0060] The technical effects and advantages of this application are as follows:

[0061] Compared with existing technologies, this method and system for monitoring safety belt decoupling during power workers' tower climbing process overcomes the limitations of traditional "post-event alarms" by using an LSTM dynamic risk prediction model. It can predict the probability of decoupling in advance, giving workers valuable time to avoid danger. Combined with an improved SVM classification model, it accurately distinguishes between three decoupling states: "normal," "suspected," and "confirmed," avoiding work interruptions or falls caused by misjudgments or missed detections. Pilot applications have reduced the accident rate during tower climbing operations, directly protecting the lives of workers.

[0062] Based on a five-level risk quantification system (extremely low-low-medium-high-extremely high) using fuzzy comprehensive evaluation, a four-level early warning mechanism (local alert - on-site push - platform linkage - rescue dispatch) is matched to ensure that risk response is "not skipped levels and not delayed". If a level four early warning is triggered, the system can automatically push the GPS location of the workers, details of the decoupling risk point, and emergency plan to the management system, and link up to initiate rescue dispatch, effectively improving emergency response efficiency and reducing the severity of accident consequences.

[0063] An improved Kalman filter algorithm (with real-time updates of observation noise covariance) is adopted to improve the signal-to-noise ratio of the data and effectively filter out dynamic interference such as wind and sand impact and tower vibration. Combined with the improved DS evidence theory that introduces sensor reliability weights, the judgment bias caused by interference from a single sensor is solved. After the fusion of multi-source data, the global confidence error is small, providing a high-purity data foundation for subsequent state classification.

[0064] By employing a process of "sliding window sampling - time-domain and frequency-domain feature extraction - normalization," 12 decoupling-sensitive features, such as the peak acceleration factor and the proportion of high-frequency tension energy, were selected. Then, using a multi-model collaborative approach of "improved DS evidence theory (data fusion) + improved SVM (state classification) + LSTM (dynamic prediction) + fuzzy comprehensive evaluation (risk quantification)," a precise analysis of the entire chain from data to conclusions was formed, which improved the accuracy compared to traditional single-model analysis.

[0065] Based on the analysis of the safety belt stress and the distribution of decoupling risk points, sensors are precisely deployed at three key locations: the main hook-safety rope, the auxiliary hook-tower fixing point, and the middle section of the safety rope. The sensors are fixed in a close fit (installation gap ≤ 0.1mm) and use anti-pull cables (tensile strength ≥ 500N) to avoid interference with normal operation by sensor installation, while ensuring complete collection of key decoupling signals, covering all risk points in the tower climbing operation process. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the method of this application;

[0067] Figure 2 This is a schematic diagram of the system flow of this application.

[0068] In the diagram: 1. Perception layer; 2. Data layer; 3. Analysis layer; 4. Application layer. Detailed Implementation

[0069] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0070] Example 1: As shown in the attached document Figure 1 The method shown is a monitoring method for preventing safety belts from becoming detached during the process of power workers climbing towers. The method includes the following steps:

[0071] S1. Decoupling Monitoring Sensor Deployment and Baseline Data Calibration: MEMS sensors are selected and installed at key risk points of the seat belt. Baseline data calibration is completed through zero-point calibration and multi-scenario simulation.

[0072] The MEMS sensors in S1 include a triaxial accelerometer, a tension sensor, and a displacement sensor. The selected sensors are a triaxial accelerometer (model ADXL355), a tension sensor (model LSZ-A), and a displacement sensor (model KTM-20) based on microelectromechanical systems (MEMS) technology. The key risk points of the safety belt include the connection between the main hook and the safety rope, the connection between the auxiliary hook and the tower anchor point, and the tension monitoring point in the middle section of the safety rope. Specifically: at the connection between the main hook and the safety rope, an accelerometer and a tension sensor are installed to monitor sudden acceleration changes and a rapid drop in tension when the hook detaches.

[0073] At the connection point between the auxiliary hook and the tower body: install displacement and acceleration sensors to monitor hook displacement and vibration.

[0074] Tension monitoring point in the middle section of the safety rope: Install a tension sensor to help verify the overall stress state of the safety rope;

[0075] The sensor sensitivity threshold is adjusted based on the decoupling mechanical characteristic curve; the sensitivity adjustment formula is as follows: ,in Sensitivity; Initial sensitivity; , These represent the maximum and minimum tensions during the unhooking process; To improve the accuracy of target tension detection.

[0076] During installation, a bonding method is adopted, using high-strength epoxy resin to tightly bond the sensor to the connection point with an installation gap of ≤0.1mm to eliminate signal acquisition deviation caused by gaps. At the same time, the sensor cable adopts an anti-pull design with a tensile strength of ≥500N to prevent cable breakage during operation.

[0077] The specific strategy for zero-point calibration in S1 is as follows: place the sensor in a static environment, collect 100 sets of static data, calculate the average value as the zero-point reference, and eliminate initial data errors; for example, the zero-point reference of a tension sensor. for: ,in For the first Group static tension value;

[0078] The specific strategy for benchmark data calibration through multi-scenario simulation is as follows: record the sensor's response data in each scenario, and establish a benchmark threshold library for three states: "normal state, suspected decoupling, and confirmed decoupling." For example, the tension threshold for confirmed decoupling is... ,in If the average tension during normal operation is 1200N, then ;

[0079] In multi-scenario simulations, three typical decoupling scenarios were simulated using a calibration device (model FLUKE754):

[0080] Scenario 1: The main hook accidentally falls off, simulating the tension dropping suddenly from 1500N to 0N, and the acceleration increasing suddenly from 0g to 5g;

[0081] Scenario 2: The secondary hook is not fully locked, and the simulated displacement offset increases from 0mm to 10mm, with an acceleration fluctuation of ±2g;

[0082] Scenario 3: The tension of the safety rope suddenly drops, simulating a decrease from 1200N to 800N over a period of 5 seconds.

[0083] S2. Decoupling monitoring data acquisition and preprocessing: Real-time acquisition of sensor physical signals, followed by outlier removal, data completion and adaptive filtering to obtain clean and valid data;

[0084] In S2, the adaptive filtering process uses an improved Kalman filter algorithm, and the filter update formula is:

[0085] State prediction: ;

[0086] Covariance prediction: ;

[0087] Kalman gain: ,in Variance is updated in real time via a sliding window. , The length of the window;

[0088] Status Update: ;

[0089] Covariance update: ;

[0090] in, for Predicted state value at any given time; for State value after filtering at any given time; for State value after filtering at any given time; This is the state transition matrix; For control matrix; To control the quantity; It is the covariance matrix; for Time-prediction error covariance matrix; for Time error covariance matrix; For process noise covariance; The observation matrix; for Time-based observations; To observe the noise covariance; It is the identity matrix; for Time-based Kalman gain; Find the inverse of a matrix; This is the matrix transpose.

[0091] By adopting the above technical solution, the sensor collects data in real time. The collection parameters are set as follows: sampling frequency: 100Hz for acceleration sensor, 50Hz for tension sensor, and 20Hz for displacement sensor.

[0092] Data collected includes: acceleration (x, y, z axes), tension, displacement, and temperature and humidity data of the sensor itself.

[0093] Data transmission: Data is transmitted to the edge computing gateway via Bluetooth 5.0 protocol.

[0094] Preliminary screening of the collected raw data:

[0095] Outlier removal: The 3σ criterion is used; Data completion: If there is missing data (i.e., the missing rate is ≤5%), linear interpolation is used to complete the data; Validity verification: Data that conforms to the fluctuation range of the seat belt under normal working conditions is retained, such as the normal range of acceleration ±1g and the normal range of displacement ±3mm, to ensure that the data used for analysis is complete and reliable.

[0096] To address environmental interference during tower climbing operations, such as wind and sand impact (high-frequency interference, 100Hz~500Hz) and tower vibration (low-frequency interference, 1Hz~10Hz), an improved Kalman filter algorithm is employed: State vector definition. ,in , , It is a triaxial acceleration. For tension, For displacement;

[0097] State transition matrix Observation matrix Process noise covariance ,in It is a diagonal matrix;

[0098] Window length After filtering, the signal-to-noise ratio of the data is improved, effectively removing the influence of environmental interference on the real signal.

[0099] S3. Multi-dimensional intelligent analysis of decoupling behavior: Extract decoupling sensitive features through feature engineering, fuse multi-sensor data based on improved DS evidence theory, use improved SVM to classify decoupling status, combine LSTM model to predict dynamic risk trends, and finally complete risk level quantification through fuzzy comprehensive evaluation.

[0100] The specific strategy for improving the DS evidence theory fusion of multi-sensor data in S3 is as follows:

[0101] Introducing sensor reliability weights ,and ,in For the first Historical misjudgment rate of each sensor Number of sensors;

[0102] Conflict evidence processing employs a weighted average adjusted BPA value, the adjustment formula being: , To identify a subset of states in the framework, For the first Each sensor pair The original BPA value, For the first Each sensor pair The corrected BPA value; the fusion formula is .

[0103] By adopting the above technical solution, and targeting 12 sensitive features (4 per channel) from the three sensors (main hook, auxiliary hook, and safety rope), multi-source data is integrated through improved DS evidence theory:

[0104] Evidence Body and Identification Framework Definition: The sensitive features of the three sensors are respectively used as evidence bodies. (Main hook) (Secondary hook) (Safety rope), identification frame ;

[0105] Sensor reliability weight calculation: Based on 100 preliminary simulation tests, the false judgment rate of each sensor was statistically analyzed. (Main hook) (Secondary hook) 6 (Safety rope), according to the weight calculation formula Calculations yielded , , ;

[0106] BPA Trust Calculation: Euclidean distance is used to measure the similarity between the current feature and the baseline feature. The distance formula is... ,in For the first The first piece of evidence One characteristic, for The state of One benchmark feature; Trust conversion formula The original BPA values ​​of each piece of evidence were obtained.

[0107] Conflicting Evidence Processing and Fusion: Calculating the Conflict Coefficient Between Evidence Bodies , For high conflict, a weighted average BPA value is used to correct it. The weighted average BPA value was corrected. To obtain global trust level , , The preliminary assessment is that it is either suspected or confirmed to have decoupled.

[0108] The specific strategy for extracting decoupling-sensitive features in S3 is as follows: Based on the preprocessed clean data, decoupling-sensitive features are selected through a process of "sliding window sampling - time-domain and frequency-domain feature extraction - feature normalization".

[0109] Sliding window sampling: Based on the previously calibrated average response time of decoupling (0.5s), the sliding window duration is set to 0.5s, and the window step size is 0.17s (1 / 3 of the window duration). The data is segmented to ensure coverage of the signal changes at the moment of decoupling.

[0110] Temporal feature extraction: Calculate the mean of the data within each window. ,variance Peak Peak factor , cliff ,in For data within the window, For window data volume;

[0111] Frequency domain feature extraction: The db4 wavelet basis function is used to perform a three-level decomposition of acceleration and tension data to separate low-frequency signals (normal operation, 0Hz~10Hz) and high-frequency signals (decoupling abrupt change, 10Hz~100Hz), and the energy proportion of the high-frequency signal is calculated. Wavelet entropy , The normalized probability of wavelet coefficients. For the number of coefficients, High-frequency signal energy refers to the energy contained in the "high-frequency signal" separated after wavelet decomposition of sensor data such as acceleration and tension. This type of high-frequency signal corresponds to the "sudden change in the action of the safety belt unhooking" during tower climbing operations (such as a sudden change in acceleration and a sudden drop in tension when the hook falls off, with a frequency usually between 10Hz and 100Hz). The total signal energy refers to the total energy of the original sensor signal before decomposition (including low-frequency signals from normal operation and high-frequency signals related to decoupling). By calculating the "proportion of high-frequency energy in the total energy," the significance of the "decoupling abrupt change characteristic" in the signal is quantified. During normal tower climbing operations, the signal is mainly low-frequency (1Hz~10Hz, such as the smooth movement of slow climbing). The proportion is low; if decoupling occurs, the high-frequency mutation signal is enhanced. The proportion will increase significantly, which will serve as one of the key characteristics for subsequent judgments on "whether there is a risk of decoupling";

[0112] Feature normalization: Min-Max normalization is used to eliminate the influence of dimensions and adjust the feature values ​​to the range of [0,1] to ensure that the weight of each feature on the analysis results is balanced.

[0113] The decoupling state classification of the improved SVM in S3 specifically involves: using the Grey Wolf Optimization Algorithm (GWO) to optimize the penalty coefficient of the SVM. With kernel function parameters Optimize the objective function for: ,in For real labels, For predicting labels, For the number of samples; SVM classification decision function for ,in For the number of support vectors, For Lagrange multipliers, It is a radial basis kernel function, and , For bias terms, It is a symbolic function.

[0114] The specific strategy for decoupling state classification and determination based on improved SVM in S3 is as follows:

[0115] An improved SVM classification model is constructed using global trust score and 12 normalized sensitive features as input:

[0116] Training set construction: Collect 1000 samples (300 normal operation samples, 300 suspected decoupling samples, and 400 confirmed decoupling samples). Each sample contains 15 input features (3 trust features + 12 sensitivity features). Output labels. Corresponding to three types of states;

[0117] Model parameter optimization: The Grey Wolf Optimization (GWO) algorithm is used to optimize the penalty coefficient of the SVM. With kernel function parameters The optimization objective was to achieve the highest accuracy with 5-fold cross-validation; the GWO algorithm iterated 20 times with an initial population size of 10, ultimately yielding the optimal parameters. and ;

[0118] Classification Decision: SVM Classification Decision Function ,in For the number of support vectors, For Lagrange multipliers, Set the probability threshold to 0.8. If the maximum probability is ≥0.8, output the judgment result; otherwise, trigger a second data collection.

[0119] The input for dynamic risk prediction based on the LSTM model in S3 is a sequence of sensitive features over the past five time windows. The "historical input data" of the LSTM model, consisting of "decoupling-sensitive features" from nearly five time windows, forms the basis for predicting future risks. Given a set of multidimensional feature vectors (e.g., 12-dimensional, including decoupling-sensitive features such as peak acceleration of the main / secondary hooks, variance of safety rope tension, and proportion of high-frequency energy), the output is the risk probability for the next two time windows. ,in For the next time window ( The probability of a definite decoupling occurring; For the next time window ( Given the probability of a definite decoupling, the state update strategy of the LSTM model is as follows:

[0120] Forgotten Gate: ;

[0121] Input Gate: ;

[0122] Cell state: ;

[0123] Output gate: ;

[0124] in, , , These are the forget gate, input gate, and output gate, respectively. , These represent the current cell state and the current hidden state, respectively. , These represent the cell state at the previous moment and the hidden state at the previous moment, respectively. , , , This is the weight matrix; , , , For bias terms; It is the sigmoid function; This is element-wise multiplication; This represents the current state of the candidate cells; It is the hyperbolic tangent function.

[0125] The LSTM model structure consists of 2 LSTM layers (64 neurons per layer) + 1 fully connected layer (32 neurons) + 1 output layer (2 neurons), with activation functions of tanh (LSTM layer), ReLU (fully connected layer), and sigmoid (output layer).

[0126] Model training: Adam optimizer was used, with a learning rate of 0.001, a batch size of 32, 100 iterations, and cross-entropy loss function. ,in For batch sample size, For true probability, To predict probabilities;

[0127] Predictive applications: If predicting This will trigger an early warning, giving workers ≥0.3 seconds to avoid danger.

[0128] The specific strategy for quantifying risk levels through fuzzy comprehensive evaluation in S3 is as follows:

[0129] Specifically: Select the displacement deviation rate Duration of acceleration exceeding the limit Tension decrease range Integrating Trust LSTM predicted risk value To evaluate the indicators, the indicator weights are determined using the analytic hierarchy process (AHP), and the consistency test formula is as follows: , , To determine the largest eigenvalue of a matrix, For random consistency index; fuzzy evaluation matrix Comprehensive evaluation results , For the indicator weight vector, according to The final risk level is determined by the fuzzy level corresponding to the maximum value in the middle.

[0130] Among them, displacement deviation rate Tension decrease range Integrating Trust LSTM predicted risk value Evaluation indicators ;

[0131] The strategy for determining indicator weights is as follows: Five power safety experts were invited to construct a judgment matrix using the 1-9 scaling method, and the weights were calculated using the analytic hierarchy process (AHP). Consistency checks were then performed. Weight vector ;

[0132] Fuzzy level classification: Each indicator is divided into 5 fuzzy levels. For example, the tension reduction level standards are: extremely low (≤10%), low (10%~20%), medium (20%~30%), high (30%~50%), and extremely high (≥50%).

[0133] Membership degree determination: The triangular membership function is used to calculate the membership degree of each indicator to the level, such as the tension decrease. Membership degree of "high" level Membership degree of "extremely high" level ;

[0134] Comprehensive evaluation: Constructing a fuzzy evaluation matrix Calculate the comprehensive evaluation results The maximum value corresponds to the "high" level, thus the risk level is determined to be high risk.

[0135] S4. Risk Level Early Warning Response for Decoupling: Based on the risk level, trigger local alarms, remote push notifications and emergency response measures to provide a level early warning system, while also enabling visualized monitoring of the work scenario based on a digital twin model.

[0136] The risk classification and early warning response strategy for decoupling is as follows:

[0137] Based on the risk level (extremely low, low, medium, high, extremely high), a four-level early warning mechanism is triggered accordingly:

[0138] Level 1 Warning (Extremely Low to Low Risk): Only minor local warnings are displayed.

[0139] Level 2 Warning (Medium Risk): Local alarm + on-site monitoring push notifications;

[0140] Level 3 Warning (High Risk): Local strong alarm + remote platform push + emergency preparedness;

[0141] Level 4 Warning (Extremely High Risk): Multi-channel alarm system + emergency response linkage + rescue dispatch.

[0142] The tiered early warning execution strategy is as follows:

[0143] Level 1 warning: The LED warning light on the safety belt will activate (flashing green, 1Hz frequency) without a buzzer, reminding the operator to check.

[0144] Level 2 Warning: The LED warning light flashes red (frequency 2Hz), the buzzer emits an intermittent warning sound (1 second sound, 1 second silence), and simultaneously pushes warning information to the on-site monitoring personnel's mobile APP via 4G network, including the location of abnormal sensors (e.g., "main hook sensor"), the current risk value (e.g., ... ) and recommended measures ("Stop work and check the main hook");

[0145] Level 3 warning: LED warning light stays red, buzzer sounds continuously (volume ≥ 85dB), pushes real-time data and 3D visualization to the monitoring center platform, the platform automatically marks the location of the workers, and activates the on-site camera to point at the risk area.

[0146] Level 4 Early Warning: Based on Level 3 early warning, the emergency linkage interface automatically connects to the power operation management system, pushing the GPS location of the operator (error ≤ 5m), details of the risk point of decoupling, and emergency response plan (such as "immediately stop climbing the tower and activate the backup safety rope"). At the same time, it triggers the rescue dispatch process, notifying the nearest rescue personnel to rush to the scene. The data transmission delay is ≤ 500ms to ensure timely emergency response.

[0147] Example 2, please refer to Figure 2 A monitoring system for preventing safety belt derailment during power workers' tower climbing process, the system comprising:

[0148] The sensing layer consists of MEMS sensors, a positioning module, and temperature and humidity sensors, and is used to collect physical signals and environmental information.

[0149] Data layer: includes edge computing gateway and cloud database. The edge computing gateway performs data preprocessing, and the cloud database stores raw data, benchmark data and analysis results.

[0150] Analysis Layer: Deploys an improved DS evidence theory fusion module, an improved SVM classification module, an LSTM prediction module, and a fuzzy comprehensive evaluation module to complete multi-dimensional intelligent analysis. Among them, the improved DS evidence theory fusion module is used to perform improved DS evidence theory fusion of multi-sensor data; the improved SVM classification module achieves decoupled state classification through improved SVM; the LSTM prediction module combines an LSTM model to predict dynamic risk trends; and the fuzzy comprehensive evaluation module completes risk level quantification through fuzzy comprehensive evaluation.

[0151] Application layer: includes local alarm terminals, monitoring center platform, and digital twin visualization module, to realize hierarchical early warning and visual monitoring.

[0152] The digital twin visualization module constructs a 3D model of the tower climbing operation scene using Unity3D, synchronizes the perception layer data in real time, and updates the model at a frequency of 10Hz. It supports personnel location tracking, visualization of safety belt connection status, and highlighting of risk areas. The emergency linkage interface can be connected to the power operation management system. When a level 3 warning is triggered, it automatically pushes the personnel location, decoupling risk point, and emergency response plan to the management system, and initiates the rescue dispatch process. The interface communication protocol uses MQTT.

[0153] The digital twin visualization monitoring strategy is as follows:

[0154] Model building: Import CAD drawings of transmission towers and BIM models of workers, add sensor location markers and data display panels, and achieve a model texture accuracy of 1024×1024 pixels;

[0155] Real-time synchronization: Data from the perception layer is acquired through the edge computing gateway to enable real-time tracking of the operator's location (with an error of ≤0.5m from GPS), color marking of the safety belt connection status (green for normal, yellow for suspected, and red for unhooked), and highlighting of risk areas (red halo covering a 1m radius around the risk point).

[0156] Interactive features: Supports monitoring personnel to zoom and rotate the scene, click on sensor markers to view real-time data curves (acceleration and tension changes in the last 10 seconds), replay the operation process in the last 5 minutes, assist in analyzing the causes of risks, and improve the intuitiveness of monitoring and the efficiency of emergency response.

[0157] Finally: The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for monitoring against disengagement of a safety belt during a tower climbing process of a power worker, characterized by, The method comprises the following steps: S1. Unhooking monitoring sensor deployment and reference data calibration: the sensor is selected as a MEMS sensor installed at a key risk point of a safety belt, and reference data calibration is completed through zero point calibration and multi-scene simulation; S2. Unhooking monitoring data acquisition and preprocessing: real-time acquisition of physical signals of the sensor, elimination of abnormal values, data completion, and adaptive filtering processing to obtain pure and effective data; S3. Multi-dimensional intelligent analysis of unhooking behavior: sensitive features of unhooking are extracted through feature engineering, multi-sensor data are fused based on improved D-S evidence theory, unhooking state classification is realized by using improved SVM, dynamic risk trend is predicted by combining an LSTM model, and risk level quantification is completed through fuzzy comprehensive evaluation; S4. Unhooking risk grading early warning response: local alarm, remote push and emergency linkage grading early warning measures are triggered according to the risk level, and operation scene visual monitoring is realized based on a digital twin model.

2. The method of claim 1, wherein: The MEMS sensor in S1 comprises a three-axis acceleration sensor, a tension sensor and a displacement sensor; the key risk points of the safety belt comprise a main hook and a safety rope connection, a secondary hook and a tower body fixed point connection, and a safety rope middle segment tension monitoring point, specifically: the main hook and the safety rope connection: an acceleration sensor and a tension sensor are installed to monitor the acceleration mutation and tension drop when the hook falls off; the secondary hook and the tower body fixed point connection: a displacement sensor and an acceleration sensor are installed to monitor the displacement offset and vibration state of the hook; the safety rope middle segment tension monitoring point: a tension sensor is installed to assist in verifying the overall stress state of the safety rope; The specific strategy for completing reference data calibration through zero point calibration in S1 is: the sensor is placed in a static environment, 100 groups of static data are collected, the mean value is calculated as the zero point reference, and the initial data error is eliminated; The specific strategy for completing reference data calibration through multi-scene simulation is: the response data of the sensor in each scene are recorded, and a reference threshold library of three types of states, i.e., normal state, suspected unhooking and determined unhooking, is established. Adjust the sensor sensitivity threshold based on the decoupling mechanical characteristic curve, and the sensitivity adjustment formula is wherein is the sensitivity; is the initial sensitivity; , are the maximum and minimum tension of the decoupling process, respectively; is the target tension detection accuracy.

3. The method of claim 2, wherein: The adaptive filtering processing in S2 adopts an improved Kalman filtering algorithm, and the filtering update formula is: The specific strategy for fusing multi-sensor data by using the improved D-S evidence theory in S3 is:

4. The method of claim 3, wherein: The specific strategy for completing risk level quantification through fuzzy comprehensive evaluation in S3 is: State prediction: ; Covariance prediction: ; Kalman gain: where updated in real-time by a sliding window variance, , is the window length; Status update: ; Covariance update: ; wherein, is the state prediction value at time instant k; is the filtered state value at time instant k; is the filtered state value at time instant k; is a state transition matrix; is a control matrix; is a control quantity; is a covariance matrix; is the prediction error covariance matrix at time instant k; is the error covariance matrix at time instant k; is a process noise covariance; is an observation matrix; is the observation value at time instant k; is an observation noise covariance; is an identity matrix; is the Kalman gain at time instant k; is a matrix inversion; is a matrix transpose.

5. The method of claim 4, wherein: The system comprises: Introducing sensor reliability weights , and wherein is the historical false rejection rate of the th sensor, is the number of sensors; Conflict evidence processing employs a weighted average adjusted BPA value, the adjustment formula being: , To identify a subset of states in the framework, For the first Each sensor pair The original BPA value, For the first Each sensor pair The corrected BPA value; the fusion formula is .

6. The method of claim 5, wherein: The improved SVM in the S3 is used for the decoupling state classification, specifically, a grey wolf optimization algorithm (GWO) is used to optimize a penalty coefficient of the SVM and a kernel function parameter , and an objective function is optimized , wherein is a true label, is a predicted label, is a sample number; a SVM classification decision function is , wherein is a support vector number, is a Lagrange multiplier, is a radial basis kernel function, and , is a bias term, is a sign function.

7. The method of claim 6, wherein: The dynamic risk prediction input based on the LSTM model in the S3 is a sensitive feature sequence of a recent 5 time windows , and the output is a risk probability of a future 2 time windows , wherein is a probability of occurrence of a certain disengagement in a next time window ; is a probability of occurrence of a certain disengagement in a next-next time window , and the state updating strategy of the LSTM model is: forget gate: ; Input gate: ; Cell state: ; Output gate: ; wherein, , , are respectively forget gate, input gate, output gate output; , are respectively current cell state and current hidden state; , are respectively previous moment cell state and previous moment hidden state; , , , is a weight matrix; , , , is a bias term; is a sigmoid function; is an element-wise multiplication; is a current candidate cell state; is a hyperbolic tangent function.

8. The method of claim 7, wherein: a perception layer composed of MEMS sensors, a positioning module and a temperature and humidity sensor, for collecting physical signals and environmental information; Specifically: select displacement deviation rate , acceleration overrun duration , tension drop amplitude , fusion trust degree , LSTM prediction risk value Evaluation index, index weight is determined by AHP, consistency check formula is , , The maximum eigenvalue of the judgment matrix is Random consistency index; fuzzy evaluation matrix , comprehensive evaluation result , The index weight vector is determined according to the maximum value of the fuzzy level corresponding to the maximum value in The final risk level.

9. A safety belt unhooking prevention monitoring system for a power worker tower climbing process, which implements the method according to any one of claims 1 to 8, characterized by, a data layer comprising an edge computing gateway and a cloud database, the edge computing gateway realizes data preprocessing, and the cloud database stores original data, reference data and analysis results; an analysis layer: an improved D-S evidence theory fusion module, an improved SVM classification module, an LSTM prediction module and a fuzzy comprehensive evaluation module are deployed to complete multi-dimensional intelligent analysis, wherein the improved D-S evidence theory fusion module is used to fuse multi-sensor data by using the improved D-S evidence theory; the improved SVM classification module realizes unhooking state classification by using the improved SVM; the LSTM prediction module predicts dynamic risk trend by combining the LSTM model; and the fuzzy comprehensive evaluation module completes risk level quantification through fuzzy comprehensive evaluation; ​ ​ Application layer: including local alarm terminal, monitoring center platform, digital twin visualization module, realizing hierarchical early warning and visual monitoring.

10. The powerline worker climb procedure safety harness unhooking prevention monitoring system of claim 9, wherein: The digital twin visualization module constructs a tower climbing operation scene three-dimensional model through Unity3D, synchronizes the perception layer data in real time, the model update frequency is 10 Hz, supports operation personnel position tracking, safety belt connection state visualization and risk area highlight display, the emergency linkage interface can be connected with the power operation management system, when triggering the third level early warning, automatically pushes the operation personnel position, unhooking risk point and emergency disposal scheme to the management system, linkage starts the rescue flow process, the interface communication protocol adopts MQTT.

Images