Intelligent monitoring system and method for aerial work safety rope

Abstract

The application belongs to the technical field of operation safety monitoring, and relates to a high-altitude operation safety rope intelligent monitoring system and method. The method collects strain signals of key stress points of the safety rope, adopts a composite noise reduction method combining a sliding window dynamic statistics and a variational mode decomposition to perform adaptive noise reduction preprocessing on the strain signals, constructs a machine learning model and performs training, adopts a weight initialization strategy based on strain extreme value distribution during the training, adopts a dynamic sparse regularization method to optimize model parameters, collects real-time strain signals, inputs the trained machine learning model to predict a predicted strain value after adaptive noise reduction preprocessing, compares the predicted strain value with a safety threshold reference, and sends an early warning signal according to the current operation height and motion state. The application improves the adaptability to complex loads and extreme working conditions, and improves the monitoring and early warning accuracy.

Description

Technical Field

[0001] This invention belongs to the field of work safety monitoring technology, and in particular relates to an intelligent monitoring system and method for safety ropes used in high-altitude operations. Background Technology

[0002] With the widespread application of high-altitude operations, the real-time monitoring of the performance and status of safety ropes, as crucial equipment for ensuring worker safety, has become paramount. Traditional high-altitude operation safety rope monitoring systems largely rely on simple sensors and static threshold judgments, making them ill-equipped to handle complex and ever-changing working environments. This is especially true under conditions of high dynamic loads, extreme working conditions, and significant strain signal noise, which can easily lead to inaccurate monitoring results and even false or missed alarms. Therefore, improving the adaptability and accuracy of monitoring systems under complex conditions, particularly their early warning capabilities under dynamic loads and abnormal states, has become a pressing technical challenge.

[0003] Chinese invention patent CN119832683A discloses a safety monitoring system for omnidirectional safety belts, including a controller, power module, environmental monitoring module, position positioning module, audible and visual alarm module, safety rope wearing monitoring module, and remote data transmission module. The safety rope wearing monitoring module includes four pressure sensors, corresponding to four D-shaped loops on the chest, back, left side, and right side of the omnidirectional safety belt. The controller detects a pressure signal from any one of these sensors and determines that the worker is in a protected state. It also includes an accelerometer installed on the omnidirectional safety belt to monitor the real-time movement of the worker and determine whether they are moving or stationary. While this improves the comprehensiveness of safety belt wearing detection, it cannot provide early warning of safety belt or rope breakage risks. Furthermore, it still uses simple threshold comparisons of discrete signals, which is insufficient for capturing the continuous strain characteristics of the safety rope under dynamic stress and struggling to handle high-frequency vibration noise and baseline shifts after long-term use, resulting in insufficient sensitivity for warning of critical breakage states.

[0004] In summary, existing technologies fail to effectively address three core issues: first, insufficient signal denoising and baseline correction capabilities under dynamic loads lead to low accuracy in monitoring data; second, incomplete capture of multi-scale strain characteristics (such as instantaneous impacts and sustained loads) easily results in missed alarms; and third, the risk of overfitting exists in model training, and the tensile-compressive asymmetry of strain signals is not considered, leading to poor early warning stability under complex working conditions. Therefore, developing a high-precision intelligent monitoring system for safety ropes that can adapt to high dynamic loads, strong noise interference, and complex working conditions is crucial to overcoming the bottlenecks of existing technologies. Summary of the Invention

[0005] This invention aims to solve the problems of poor adaptability, low monitoring accuracy, and insufficient early warning reliability of existing high-altitude work safety rope monitoring systems under complex working conditions, and provides an intelligent monitoring system and method for high-altitude work safety ropes.

[0006] In a first aspect, the present invention provides an intelligent monitoring method for safety ropes used in high-altitude operations, comprising:

[0007] Collect strain signals at key stress points of the safety rope and simultaneously acquire working height and motion status;

[0008] An adaptive noise reduction preprocessing method combining sliding window dynamic statistics and variational mode decomposition is used to denoise the strain signal.

[0009] A machine learning model is constructed and trained. During training, a weight initialization strategy based on strain extreme value distribution is adopted, and a dynamic sparse regularization method is used to optimize the model parameters.

[0010] The machine learning model includes a multi-scale temporal feature fusion module composed of multiple convolutional layers and an improved GLU gating unit, and a load-level attention mechanism. The denoised strain signal is processed by convolutional layers with different kernel sizes to obtain convolutional features of different scales. The convolutional features of different scales are then enhanced by the improved GLU gating unit. The channels of the convolutional features of different scales are concatenated to obtain multi-scale fused features. The load-level attention mechanism calculates local attention weights and global attention weights for the multi-scale fused features, and then sums them by weight and concatenates them with the output features processed by the improved GLU gating unit to obtain the final attention output. The final attention output is then passed through a linear activation function to obtain the predicted strain value.

[0011] Real-time strain signals are collected, and after adaptive noise reduction preprocessing, they are input into a trained machine learning model to predict strain values. The safety threshold benchmark is dynamically adjusted according to the current working height and motion state. The predicted strain values ​​are compared with the safety threshold benchmark, and an early warning signal is issued.

[0012] Further preferably, the adaptive noise reduction preprocessing includes: first, dividing the original strain signal into multiple sliding windows, with the data within each sliding window considered as an independent processing unit; then, applying variational mode decomposition to the data within each sliding window to decompose it into multiple modal components; then, calculating the corresponding adaptive frequency band weights based on the difference between the center frequency and the load characteristic frequency reference value of each modal component; and recombinating the modal components by weighted summation to obtain the preliminary noise-reduced signal; finally, smoothing the preliminary noise-reduced signal using sliding window mean filtering and further optimizing the preliminary noise-reduced signal by combining it with a noise suppression coefficient to obtain the noise-reduced strain signal.

[0013] Further preferably, the adaptive noise reduction preprocessing is expressed as:

[0014] ;

[0015] In the formula, The strain signal after noise reduction at time t within the w-th sliding window; This represents the original strain signal at time t within the w-th sliding window; The number of variational mode decomposition layers; The adaptive frequency band weight is calculated as follows: ; For the first The center frequency of each modal component; The load characteristic frequency reference value is calculated as follows: ; Let be the set of frequencies corresponding to the significant harmonic components extracted from the original strain signal by Fast Fourier Transform, and max be the function for maximizing the harmonic components. This is the penalty factor for variational mode decomposition; For the first One modal component; The noise suppression coefficient is calculated as follows: ; The residuals are calculated through empirical mode decomposition. The variance of the original strain signal; This represents a sliding window mean filter; It is a time variable; This represents the length of the time window.

[0016] Further optimization reveals that the processing procedure of the multi-scale temporal feature fusion module is as follows:

[0017] ;

[0018] The first input feature is equal to ;

[0019] The second input feature is equal to ;

[0020] For the third input feature, equal to ;

[0021] in, The output characteristics are those processed by the improved GLU gating unit; For gating weights; For the Sigmoid function; Element-wise multiplication; The convolution weight matrix; The strain signal after noise reduction at time t; For depthwise separable convolutions with a kernel size of 3; For depthwise separable convolutions with a kernel size of 5; It is a depthwise separable convolution with a kernel size of 7; AvgPool is average pooling, and MaxPool is max pooling.

[0022] Further preferred, the processing procedure of the load-level attention mechanism is as follows: First, the multi-scale fusion features are divided into local time windows, and local time window features are extracted. Then, for the local time window features, local attention weights are obtained through the interaction calculation of the query matrix and the key matrix, which are used to capture the feature associations within the local time window. Global attention weights are obtained by weighting the multi-scale fusion features through the global weight matrix, which are used to capture the importance of global features. Finally, the local attention weights and global attention weights are respectively weighted and summed with the multi-scale fusion features, and the weighted summation result is concatenated with the output features processed by the improved GLU gating unit to obtain the final attention output.

[0023] Further preferably, the weight initialization strategy based on strain extreme value distribution is expressed as:

[0024] ;

[0025] ;

[0026] In the formula, To initialize the weights, specifically all the weight parameters of the machine learning model; The mean is variance is The normal distribution; To initialize the mean of the weights; To initialize the variance of the weights; This represents the average tensile strain. This represents the average compressive strain. The peak variability variance is calculated as follows: ; The total number of samples; For the first The strain values ​​of each sample; This represents the peak mean.

[0027] Further optimization of the dynamic sparsity regularization method involves: employing channel-level dynamic sparsity constraints to optimize the allocation of computational resources by dynamically adjusting the sparsity intensity of each layer. First, the sparsity intensity of each layer is calculated, and this intensity is proportional to the computational cost of that layer. Then, the channel mixing norm is applied to the weights of each layer to calculate the sparsity loss. The channel mixing norm is calculated by first determining the L1 norm of the weights for each channel, and then determining the L2 norm of all channels' L1 norms, thereby achieving sparsity constraints on the weights. Finally, the sparsity losses of all layers are summed to obtain the total sparsity loss, which is then added to the total loss function of the machine learning model. By dynamically adjusting the sparsity intensity, adaptive adjustments to the sparsity constraint intensity of different layers are achieved, thereby optimizing the allocation of computational resources.

[0028] Further optimization yields the following total loss function:

[0029] ;

[0030] In the formula, Total loss; For classification loss weights; For classification loss; For regression loss; The loss is sparsity loss; the classification loss uses an improved FocalLoss, which is calculated as follows: In the formula, Total number of categories; For the c-th category; For category The true label; For category The predicted probability; This is the dynamic focusing coefficient, used to balance the loss contribution ratio between easy and difficult samples. With training rounds Linear increase, the calculation method is expressed as: ; Based on sparsity; To adjust the slope of the factor; For training rounds.

[0031] Secondly, the present invention provides an intelligent monitoring system for safety ropes used in high-altitude operations, comprising:

[0032] The data acquisition module includes a strain sensor array for acquiring strain signals at key stress points of the safety rope, as well as a temperature and humidity sensor, an altimeter, and a motion acceleration sensor.

[0033] The preprocessing module is used to perform adaptive noise reduction preprocessing on the strain signal;

[0034] The monitoring and computing module deploys a trained machine learning model and designs a safety threshold. The machine learning model includes a multi-scale temporal feature fusion module composed of multiple convolutional layers and an improved GLU gating unit, and a load-level attention mechanism. The denoised strain signal is processed by convolutional layers with different kernel sizes to obtain convolutional features of different scales. The convolutional features of different scales are respectively enhanced by the improved GLU gating unit. The channels of the convolutional features of different scales are concatenated to obtain multi-scale fused features. The load-level attention mechanism calculates local attention weights and global attention weights for the multi-scale fused features, and then sums them by weight and concatenates them with the output features processed by the improved GLU gating unit to obtain the final attention output. The final attention output is used to obtain the predicted strain value through a linear activation function.

[0035] The early warning terminal dynamically adjusts the safety threshold benchmark based on the current working height and movement status, compares the predicted strain value with the safety threshold benchmark, and issues an early warning signal.

[0036] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores computer-executable instructions, and the processor invokes the computer-executable instructions to execute the various steps of the intelligent monitoring method for safety ropes in high-altitude operations.

[0037] The beneficial effects of this invention are:

[0038] 1. By combining sliding window dynamic statistics with variational mode decomposition, noise is effectively removed and dynamic load signals are optimized, overcoming the sensitivity of traditional methods to baseline drift and ensuring the accuracy and reliability of monitoring data;

[0039] 2. By employing depthwise separable convolution to extract local features at different scales and combining it with a load-level attention mechanism, the system accurately captures multi-scale correlations and dangerous features under dynamic loads, thereby improving its adaptability to complex loads and extreme operating conditions.

[0040] 3. Channel-level dynamic sparsity constraints are adopted to optimize the allocation of computing resources and reduce overfitting by adaptively adjusting the sparsity intensity of each layer.

[0041] 4. By modifying the initialization strategy for strain extreme value distribution, the initial performance of the model is optimized, thereby improving training efficiency and accuracy.

[0042] 5. Dynamic adjustment of safety thresholds and self-calibration technology ensure the safety of high-altitude operations, provide timely warnings, and automatically generate maintenance suggestions. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the system modules of the present invention.

[0044] Figure 2 This is a flowchart of the present invention.

[0045] Figure 3 This is a schematic diagram of a machine learning model.

[0046] Figure 4 The bar chart shows the signal-to-noise ratio for different noise reduction methods.

[0047] Figure 5 The bar chart shows the mean square error of different noise reduction methods.

[0048] Figure 6 The bar chart shows the computation time for different noise reduction methods.

[0049] Figure 7 A comparison chart showing the classification accuracy of different models in monitoring scenarios with different risk levels.

[0050] Figure 8 This is a comparison chart of early warning accuracy at different operating heights.

[0051] Figure 9 A comparison chart of strain prediction errors under different load conditions. Detailed Implementation

[0052] The present invention will be further explained in detail below with reference to the accompanying drawings and embodiments.

[0053] Example 1

[0054] like Figure 1 As shown in this embodiment, a high-altitude work safety rope intelligent monitoring system includes:

[0055] The data acquisition module includes a strain sensor array for acquiring strain signals at key stress points of the safety rope, as well as a temperature and humidity sensor, an altimeter, and a motion acceleration sensor.

[0056] The preprocessing module is used to perform adaptive noise reduction preprocessing on the strain signal;

[0057] The monitoring and computing module deploys a trained machine learning model and designs a safety threshold. The machine learning model includes a multi-scale temporal feature fusion module composed of multiple convolutional layers and an improved GLU gating unit, and a load-level attention mechanism. The denoised strain signal is processed by convolutional layers with different kernel sizes to obtain convolutional features of different scales. The convolutional features of different scales are respectively enhanced by the improved GLU gating unit. The channels of the convolutional features of different scales are concatenated to obtain multi-scale fused features. The load-level attention mechanism calculates local attention weights and global attention weights for the multi-scale fused features, and then sums them by weight and concatenates them with the output features processed by the improved GLU gating unit to obtain the final attention output. The final attention output is used to obtain the predicted strain value through a linear activation function.

[0058] The early warning terminal dynamically adjusts the safety threshold benchmark based on the current working height and movement status, compares the predicted strain value with the safety threshold benchmark, and issues an early warning signal.

[0059] Example 2

[0060] like Figure 2 As shown, a method for intelligent monitoring of safety ropes for high-altitude operations includes the following steps:

[0061] S1. Data Acquisition: Multi-axis strain signals are acquired in real time through a strain sensor array installed at key stress points of the safety rope. Environmental temperature and humidity, working height and motion acceleration are acquired simultaneously. The sampling frequency is set to 200Hz to ensure complete capture of dynamic load characteristics. All data is transmitted to the edge computing node via Bluetooth 5.0 to form a time-aligned raw dataset.

[0062] S2. Perform adaptive noise reduction preprocessing on the strain signal.

[0063] Strain sensor data exhibits high-frequency noise interference and dynamic load fluctuation characteristics. Conventional wavelet threshold denoising methods are sensitive to baseline drift under unsteady conditions. This embodiment employs a composite denoising method combining sliding window dynamic statistics and variational mode decomposition. First, the original strain signal is divided into multiple sliding windows, with the data within each window considered an independent processing unit. Then, variational mode decomposition is applied to the data within each sliding window, decomposing it into multiple modal components. Next, based on the difference between the center frequency of each modal component and the load characteristic frequency reference value, the corresponding adaptive frequency band weight is calculated. By weighted summation, the modal components are recombined to obtain the preliminary denoised signal. Finally, the preliminary denoised signal is smoothed using sliding window mean filtering and further optimized by incorporating a noise suppression coefficient, thereby achieving effective denoising of the original strain signal, expressed as:

[0064] ;

[0065] In the formula, The strain signal after noise reduction at time t within the w-th sliding window; This represents the original strain signal at time t within the w-th sliding window; The number of variational mode decomposition layers; The adaptive frequency band weight is calculated as follows: ; For the first The center frequency of each modal component; The load characteristic frequency reference value is calculated as follows: ; Let be the set of frequencies corresponding to the significant harmonic components extracted from the original strain signal by Fast Fourier Transform, and max be the function for maximizing the harmonic components. This is the penalty factor for variational mode decomposition. Its function is to balance the sparsity and reconstruction error of mode decomposition. By adjusting this parameter, the bandwidth constraint strength of the mode components during the decomposition process can be controlled to avoid over-decomposition or under-decomposition. For the first One modal component; The noise suppression coefficient is calculated as follows: ; The residuals are calculated through empirical mode decomposition. The variance of the original strain signal; This represents a sliding window mean filter; It is a time variable; This represents the length of the time window.

[0066] In one embodiment, the superiority of the proposed composite denoising method is verified by comparing the performance of different denoising methods in terms of signal-to-noise ratio, mean square error, and computation time. The composite denoising method combining sliding window dynamic statistics and variational mode decomposition in this embodiment is compared with traditional wavelet thresholding, moving average, EMD decomposition, and Kalman filtering. The experimental results show that the method of this invention significantly improves signal quality while maintaining reasonable computational efficiency. Figures 4-6 The invention visually demonstrates the leading advantages of the composite noise reduction method in various indicators, especially the signal-to-noise ratio, which is significantly higher than other methods, while the computation time is much lower than that of methods with high computational complexity, reflecting the balanced consideration of the method in engineering practicality.

[0067] S3, Machine Learning Model Training.

[0068] like Figure 3As shown, the machine learning model includes a multi-scale temporal feature fusion module composed of multiple convolutional layers and an improved GLU gating unit, and a load-level attention mechanism. The denoised strain signal is processed by convolutional layers with different kernel sizes to obtain convolutional features of different scales. The convolutional features of different scales are respectively enhanced by the improved GLU gating unit. The channels of the convolutional features of different scales are concatenated to obtain multi-scale fused features. The load-level attention mechanism calculates the local attention weights and global attention weights of the multi-scale fused features, and then sums them by weight and concatenates them with the output features processed by the improved GLU gating unit to obtain the final attention output. The final attention output is used to obtain the predicted strain value through a linear activation function.

[0069] S301. Construct a multi-scale temporal feature fusion module

[0070] Conventional methods use fixed time windows to extract statistical features, which are difficult to capture multi-scale correlations under dynamic loads. This invention employs a hierarchical temporal convolution structure with depthwise separable convolution layers. First, depthwise separable convolution operations with different kernel sizes are applied to the denoised strain signal to extract local features at different scales. Depthwise separable convolution reduces computation and the number of parameters by decomposing the standard convolution into channel-wise convolution and pointwise convolution. Then, max pooling and average pooling operations are performed on the convolution results with kernel sizes of 5 and 7, respectively, to extract more representative features. Finally, these pooled features are fused through channel concatenation to obtain multi-scale fused features, which can better capture multi-scale correlations under dynamic loads, as shown below:

[0071] ;

[0072] In the formula, It is a multi-scale fusion feature; The strain signal after noise reduction at time t; For depthwise separable convolutions with a kernel size of 3; For depthwise separable convolutions with a kernel size of 5; For depthwise separable convolutions with a kernel size of 7; For channel concatenation operations, AvgPool is for average pooling and MaxPool is for max pooling.

[0073] Furthermore, each convolutional layer is followed by an improved GLU gating unit. For the output features of each convolutional layer, they are first input into the improved GLU gating unit. The gating unit weights the input features with gating weights to obtain a gating signal. At the same time, the input features are also convolved by the convolution weight matrix to obtain a convolutional feature. Then, the gating signal is activated by the Sigmoid function to restrict its value range to (0,1). Finally, the activated gating signal is multiplied element-wise with the convolutional feature to achieve dynamic weighting of the convolutional feature, thereby enhancing the model's ability to express important features and achieving computational lightweighting by compressing the channel dimension.

[0074] Specifically, the output characteristics after processing by the improved GLU gating unit are defined as follows: ,but, ;

[0075] The first input feature is equal to ;

[0076] The second input feature is equal to ;

[0077] For the third input feature, equal to ;

[0078] For gating weights; For the Sigmoid function; To achieve element-wise multiplication, computation is lightweighted by compressing the channel dimension; is the convolution weight matrix.

[0079] It should be noted that a conventional GLU gating unit is a single-path system, such as... ,or ,or This invention extends the traditional single-path, multiplicative output GLU into a multi-path parallel processing, additive fusion output structure. The addition operation allows features at different scales to complement and superimpose, rather than being mutually constrained as in multiplication. This is equivalent to constructing a wider and richer feature representation, which can integrate fine-grained local details and coarse-grained global context. This enables the gating unit to simultaneously perceive and fuse information from different time scales, thereby gaining a more comprehensive understanding of the dynamic behavior of strain signals, such as instantaneous impact, continuous load, and slow drift.

[0080] S302, Develop a load-level attention mechanism

[0081] To distinguish between normal operation and critical fracture state, the multi-scale fused features are first divided into local time windows, and local time window features are extracted. Then, for the local time window features, local attention weights are obtained through the interaction of the query matrix and the key matrix to capture feature correlations within the local time window. Global attention weights are obtained by weighting the multi-scale fused features using the global weight matrix to capture the importance of global features. Finally, the local attention weights and global attention weights are weighted and summed with the multi-scale fused features respectively, and the weighted sum is concatenated with the output features processed by the improved GLU gating unit to obtain the final attention output, represented as:

[0082] ;

[0083] ;

[0084] Furthermore, based on local attention weights, the multi-scale fused features and the output features processed by the GLU gated unit are fused to obtain the final attention output, expressed as:

[0085] ;

[0086] In the formula, For local attention weights; The query matrix is ​​calculated as follows: ; To query the weight matrix; This is a feature of a local time window; For time; The length of the local time window; The key matrix is ​​calculated as follows: ; This is the key weight matrix; This is the scaling factor; For global attention weights; This is the global weight matrix; For the final attention output; This is for channel splicing operations.

[0087] S303, Weight initialization strategy based on strain extreme value distribution

[0088] Traditional initialization does not consider the asymmetric tensile and compressive characteristics of the strain signal; the initialization weights of this invention follow the mean value. variance is The normal distribution is first calculated by taking the mean tensile strain and the mean compressive strain, as well as the peak variance. Then, the mean of the initial weights is calculated based on these statistics. A normal distribution is then constructed using the mean of the initial weights and the variance of the preset initial weights. Initial weights are then randomly sampled from this distribution. This method allows the initial weights to better adapt to the characteristics of the actual load direction, thereby improving the performance of the machine learning model in the early stages of training. This is expressed as:

[0089] ;

[0090] ;

[0091] In the formula, To initialize the weights, specifically all the weight parameters of the machine learning model; The mean is variance is The normal distribution; To initialize the mean of the weights; To initialize the variance of the weights; This represents the average tensile strain. This represents the average compressive strain. The peak variability variance is calculated as follows: ; The total number of samples; For the first The strain values ​​of each sample; This represents the peak mean.

[0092] S304. Employing a dynamic sparse regularization method.

[0093] To address the overfitting problem of lightweight models, a channel-level dynamic sparsity constraint is employed. This optimizes computational resource allocation by dynamically adjusting the sparsity intensity of each layer. First, the sparsity intensity of each layer is calculated, as it is proportional to the computational cost of that layer. Then, the channel mixture norm is applied to the weights of each layer to calculate its sparsity loss. The channel mixture norm is calculated by first taking the L1 norm of the weights for each channel, and then taking the L2 norm of all channels' L1 norms, thus achieving sparsity constraints on the weights. Finally, the sparsity losses of all layers are summed to obtain the total sparsity loss, which is then added to the overall loss function of the machine learning model. By dynamically adjusting the sparsity intensity, the sparsity constraint intensity of different layers is adaptively adjusted, thereby optimizing the allocation of computational resources. The calculation method is expressed as follows:

[0094] ;

[0095] ;

[0096] In the formula, For sparsity loss; This represents the total number of network layers. For layer index; For the first Layer sparsity strength; The channel mixing norm is calculated as follows: ; For the first Number of channels per layer; For the first Number of convolution kernels per layer; For the first Channel 1 Each weight; For the first Layer computational load; Based on the sparsity rate.

[0097] S305, Loss Function Optimization

[0098] To address the scarcity of safety rope breakage samples, the total loss function is first defined as a weighted sum of classification loss, regression loss, and sparsity loss. The classification loss measures the accuracy of the model in classifying the safety rope's state, the regression loss measures the accuracy of the model in predicting the safety rope's strain value, and the sparsity loss is used for sparsity regularization. Then, to further address the scarcity of safety rope breakage samples, an improved FocalLoss is applied to the classification loss. FocalLoss, by incorporating the base sparsity rate, assigns higher weights to hard-to-classify samples, thereby enhancing the model's learning ability for minority class samples. The base sparsity rate increases linearly with training epochs, allowing the machine learning model to gradually increase its focus on hard-to-classify samples during training. The calculation method is expressed as follows:

[0099] ;

[0100] In the formula, Total loss; For classification loss weights; For classification loss; For regression loss; The loss is sparsity loss; the classification loss uses an improved FocalLoss, which is calculated as follows: In the formula, Total number of categories; For the c-th category; For category The true label; For category The predicted probability; This is the dynamic focusing coefficient, used to balance the loss contribution ratio between easy and difficult samples. With training rounds Linear increase, the calculation method is expressed as: ; Based on the sparsity, preferably, ; To adjust the slope of the factor, preferably, ; For training rounds.

[0101] S306. Repeat the model training iteratively until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0102] S4. Intelligent monitoring of safety ropes for high-altitude operations.

[0103] First, the real-time sensor data is input into the preprocessing module to complete the composite noise reduction of the strain signal and the normalization of the working parameters. Then, the multi-scale time series feature fusion module extracts the multi-scale fusion features and the output features of the load graded attention mechanism. After calculation by the preset linear activation function, the predicted strain value is obtained.

[0104] Then, the system dynamically adjusts the safety threshold benchmark based on parameters such as the current working height and motion status. When the predicted strain value exceeds 80% of the critical value, a level three warning is triggered. At the same time, the edge nodes continuously upload compressed feature data to the cloud for incremental model updates.

[0105] One implementation method for dynamically adjusting the safety threshold benchmark is as follows: The system sets a basic safety threshold, but this threshold is not fixed. It is dynamically adjusted in real time based on the current working height and the movement status of the personnel. Specifically, when the working height increases, the system automatically lowers the safety threshold benchmark, making the warning more sensitive, because the higher the height, the greater the potential risk of falling, requiring earlier and more cautious warnings. At the same time, the system monitors the movement status of the personnel through acceleration sensors. If it detects situations such as rapid movement, rapid descent, or violent shaking that increase the dynamic load, it will also lower the safety threshold accordingly, thus enabling more sensitive detection of risks when the force changes drastically. Based on this, by dynamically adjusting the safety threshold benchmark, the system can automatically increase vigilance in high-risk environments, while avoiding excessive interference in stable working conditions, achieving intelligent adaptive risk warning.

[0106] Finally, the visual interface displays the stress state spectrum and health score of the safety rope in real time, and historical data can be backtracked and analyzed. When abnormal vibration patterns or progressive damage are detected, maintenance suggestions are automatically generated and pushed to the management personnel's terminal.

[0107] The system performs a self-calibration process every 8 hours, verifying sensor accuracy through a built-in calibration device to ensure the reliability of monitoring results. All alarm events are recorded with detailed timestamps, environmental parameters, and load spectrum characteristics, forming a complete traceability file.

[0108] In one embodiment, the classification accuracy of the machine learning model of this invention was compared with that of traditional methods such as Support Vector Machine (SVM), Random Forest, Long Short-Term Memory Network (LSTM), and One-Dimensional Convolutional Neural Network (1D-CNN) for monitoring scenarios with different risk levels (normal operation, minor damage, critical state, and fracture risk). Figure 7 As shown, the experimental results demonstrate that the present invention maintains stable high performance in various monitoring scenarios. In particular, in high-risk critical state and fracture risk monitoring scenarios, the accuracy decrease is significantly less than that of other methods, indicating that the load-level attention mechanism in the machine learning model of the present invention has the ability to accurately capture dangerous features.

[0109] In this embodiment, the impact of changes in operating altitude on the performance of the monitoring system is analyzed, through methods such as... Figure 8 The line graph shown illustrates the changes in warning accuracy of the three methods at different altitudes. As altitude increases, the traditional threshold method uses a fixed threshold setting, while the dynamic threshold method dynamically adjusts the safety threshold based on real-time operating height and personnel acceleration. Using a base safety threshold (10kN) as a baseline, the threshold is increased by 0.5% for every 10 meters increase in altitude. Both the traditional and dynamic threshold methods show a significant performance decrease, while the present invention, employing a dynamic threshold method, maintains a consistently high warning accuracy. The warning accuracy curve of the present invention is also flatter, verifying the system's adaptability to changes in environmental parameters.

[0110] In this embodiment, the stability of the present invention under multiple working conditions is verified by comparing the strain prediction errors under different load levels. Figure 9 The performance of four models (this invention, LSTM, 1D-CNN, and SVR) across five load ranges is shown. As the load increases, the error growth rate of the machine learning model of this invention is significantly smaller than that of the other models. Figure 9 The text specifically adds annotations in the high-load range to emphasize the advantages of the machine learning model of this invention under extreme conditions. It shows that the multi-scale feature extraction module's comprehensive capture of load features and the intelligent control of model complexity by dynamic sparse regularization enable the system to maintain accurate predictions under various load conditions.

[0111] Example 3

[0112] An electronic device includes a memory and a processor, the memory storing computer-executable instructions, the processor invoking the computer-executable instructions to perform the various steps of the intelligent monitoring method for safety ropes in high-altitude operations described in Embodiment 2.

[0113] The above description merely illustrates preferred embodiments of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make modifications or alterations to the above-disclosed content to create equivalent embodiments. However, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent monitoring of safety ropes used in high-altitude operations, characterized in that, include: The strain signals of each key stress point of the safety rope are collected by a strain sensor array, the working height is obtained synchronously by an altimeter, and the motion status is obtained by a motion acceleration sensor. A composite denoising method combining sliding window dynamic statistics and variational mode decomposition is used to perform adaptive denoising preprocessing on strain signals. This includes sliding window division, variational mode decomposition, and calculation of adaptive frequency band weights corresponding to modal components for weighting. A machine learning model is constructed and trained. During training, a weight initialization strategy based on strain extreme value distribution is adopted, and a dynamic sparse regularization method is used to optimize the model parameters. The total sparse loss is calculated and added to the total loss function of the machine learning model. By dynamically adjusting the sparse intensity, the adaptive adjustment of the sparse constraint intensity of different layers is achieved. The machine learning model includes a multi-scale temporal feature fusion module composed of multiple convolutional layers and an improved GLU gating unit, and a load-level attention mechanism. The denoised strain signal is processed by convolutional layers with different kernel sizes to obtain convolutional features of different scales. The convolutional features of different scales are respectively enhanced by the improved GLU gating unit, and then the convolutional features of different scales are concatenated to obtain multi-scale fused features. The load-level attention mechanism calculates local attention weights and global attention weights for the multi-scale fused features, and then sums them by weight and concatenates them with the output features processed by the improved GLU gating unit to obtain the final attention output. The final attention output is then passed through a linear activation function to obtain the predicted strain value. Real-time strain signals are collected, and after adaptive noise reduction preprocessing, they are input into a trained machine learning model to predict strain values. The safety threshold benchmark is dynamically adjusted according to the current working height and motion state, and the predicted strain value is compared with the safety threshold benchmark to issue an early warning signal. The processing procedure of the multi-scale temporal feature fusion module is as follows: ; The first input feature is equal to ; The second input feature is equal to ; For the third input feature, equal to ; Among them, GLU out The output characteristics are those processed by the improved GLU gating unit; For gating weights; For the Sigmoid function; For element-wise multiplication; The convolution weight matrix; The denoised strain signal at time t; DSConv 3×1 For depthwise separable convolutions with a kernel size of 3; DSConv 5×1 For depthwise separable convolutions with a kernel size of 5; DSConv 7×1 It is a depthwise separable convolution with a kernel size of 7; AvgPool is average pooling, and MaxPool is max pooling.

2. The intelligent monitoring method for safety ropes in high-altitude operations according to claim 1, characterized in that, The adaptive noise reduction preprocessing includes: The original strain signal is divided into multiple sliding windows, and the data in each sliding window is regarded as an independent processing unit. Variational mode decomposition is applied to the data within each sliding window to decompose it into multiple modal components. Then, based on the difference between the center frequency of each modal component and the load characteristic frequency reference value, the corresponding adaptive frequency band weight is calculated. Finally, the modal components are recombined by weighted summation to obtain the initial noise reduction signal. The initial denoised signal is smoothed by using a sliding window mean filter, and then optimized by combining the noise suppression coefficient to obtain the denoised strain signal.

3. The intelligent monitoring method for safety ropes in high-altitude operations according to claim 1, characterized in that, The adaptive noise reduction preprocessing is expressed as follows: ； In the formula, The strain signal after noise reduction at time t within the w-th sliding window; This represents the original strain signal at time t within the w-th sliding window; The number of variational mode decomposition layers; The adaptive frequency band weight is calculated as follows: ; For the first The center frequency of each modal component; The load characteristic frequency reference value is calculated as follows: ; Let be the set of frequencies corresponding to the significant harmonic components extracted from the original strain signal by Fast Fourier Transform, and max be the function for maximizing the harmonic components. This is the penalty factor for variational mode decomposition; For the first One modal component; The noise suppression coefficient is calculated as follows: ; For noise variance; The variance of the original strain signal; This represents a sliding window mean filter; It is a time variable; This represents the length of the time window.

4. The intelligent monitoring method for safety ropes in high-altitude operations according to claim 1, characterized in that, The processing procedure of the load-level attention mechanism is as follows: Divide the multi-scale fusion features into local time windows; For local time window features, local attention weights are obtained by interactive calculation of query matrix and key matrix, and global attention weights are obtained by weighting multi-scale fused features using global weight matrix. The local attention weights and global attention weights are weighted and summed with the multi-scale fusion features, and the weighted summation result is concatenated with the output features processed by the improved GLU gating unit to obtain the final attention output.

5. The intelligent monitoring method for safety ropes used in high-altitude operations according to claim 1, characterized in that, The weight initialization strategy based on strain extreme value distribution is expressed as follows: ； ； In the formula, To initialize the weights, specifically all the weight parameters of the machine learning model; The mean is variance is The normal distribution; To initialize the mean of the weights; To initialize the variance of the weights; This represents the average tensile strain. This represents the average compressive strain. The peak variability variance is calculated as follows: ; The total number of samples; For the first The strain values ​​of each sample; This represents the peak mean.

6. The intelligent monitoring method for safety ropes used in high-altitude operations according to claim 1, characterized in that, The process of calculating the total sparsity loss is as follows: calculate the sparsity intensity of each layer, and the sparsity intensity is proportional to the computational cost of that layer; The channel mixing norm is applied to the weights of each layer to calculate the sparsity loss of that layer. The channel mixing norm is achieved by first calculating the L1 norm of the weights of each channel, and then calculating the L2 norm of the L1 norm of all channels, thereby achieving sparsity constraints on the weights. The total sparsity loss is obtained by summing the sparse losses of all layers.

7. The intelligent monitoring method for safety ropes in high-altitude operations according to claim 6, characterized in that, The total loss function is expressed as: ； In the formula, Total loss; For classification loss weights; For classification loss; For regression loss; For sparsity loss; The classification loss uses a modified FocalLoss, and is calculated as follows: In the formula, Total number of categories; For the c-th category; For category The true label; For category The predicted probability; This is the dynamic focusing coefficient, used to balance the loss contribution ratio between easy and difficult samples. With training rounds Linear increase, the calculation method is expressed as: ; Based on sparsity; To adjust the slope of the factor; For training rounds.

8. A high-altitude work safety rope intelligent monitoring system for implementing the high-altitude work safety rope intelligent monitoring method as described in any one of claims 1-7, characterized in that, include: The data acquisition module includes a strain sensor array, a temperature and humidity sensor, an altimeter, and a motion acceleration sensor for acquiring strain signals at key stress points of the safety rope. The preprocessing module employs a composite noise reduction method combining sliding window dynamic statistics and variational mode decomposition to perform adaptive noise reduction preprocessing on the strain signal; The monitoring and computing module deploys a trained machine learning model and designs a safety threshold. The machine learning model includes a multi-scale temporal feature fusion module composed of multiple convolutional layers and an improved GLU gating unit, and a load-level attention mechanism. The denoised strain signal is processed by convolutional layers with different kernel sizes to obtain convolutional features of different scales. The convolutional features of different scales are respectively enhanced by the improved GLU gating unit. The channels of the convolutional features of different scales are concatenated to obtain multi-scale fused features. The load-level attention mechanism calculates local attention weights and global attention weights for the multi-scale fused features, and then sums them by weight and concatenates them with the output features processed by the improved GLU gating unit to obtain the final attention output. The final attention output is used to obtain the predicted strain value through a linear activation function. The early warning terminal dynamically adjusts the safety threshold benchmark based on the current working height and motion status, and compares the predicted strain value with the safety threshold benchmark to issue an early warning signal.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores computer-executable instructions, and the processor invokes the computer-executable instructions to perform the various steps of the intelligent monitoring method for safety ropes in high-altitude operations as described in any one of claims 1-7.

Images