Safety rope operation state monitoring method and system based on double-scale differential entropy features

Abstract

The application discloses a kind of safety rope operation state monitoring method and system based on double-scale differential entropy feature, it is related to safety rope monitoring technical field.Method includes collecting original stress time series data and marking, obtains stress data set with time series marking;High-frequency impact component of original stress time series data with time series marking is extracted using dynamic narrow window Gabor transform, low-frequency entropy value is calculated using fixed wide window integration combined with differential entropy operator, high-frequency impact component and low-frequency entropy value are spliced, and double-scale differential entropy feature is constructed;Neural network model is constructed and trained by double-scale differential entropy feature;The trained neural network model is deployed in safety rope monitoring terminal to carry out aerial work safety rope real-time monitoring.The application significantly improves the accuracy and robustness of aerial work safety rope state identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of safety rope monitoring, in particular to a safety rope operation state monitoring method and system based on double-scale differential entropy features. BACKGROUND

[0002] High-altitude operation is widely used in power, construction, communication, wind energy and other industries. Its operation environment is high-risk, irreversible and has very low fault tolerance. Falling accidents are one of the main causes of serious personal injury and death. Although safety ropes and other basic protective equipment are currently widely equipped, these traditional devices mostly rely on passive mechanical structures and cannot achieve intelligent identification and early warning of the operation state. Existing safety monitoring systems generally have strong lag, high false alarm rate and cannot distinguish different risk levels in terms of abnormal identification. In particular, under different working conditions such as rapid displacement of personnel, tool collision and real falling, it is often impossible to accurately identify the dangerous state due to signal aliasing or feature ambiguity. In addition, the safety rope itself carries key mechanical signals and is a direct channel for reflecting changes in the operation state. However, traditional methods cannot accurately capture the stress dynamic characteristics of the safety rope, and cannot achieve real-time and high-sensitivity analysis and judgment. Therefore, there is an urgent need for a technical system that combines high-precision sensing, intelligent feature extraction and deep learning classification to actively, real-time and accurately monitor the safety rope state during high-altitude operation, improve the operation safety protection capability and prevent accidents.

[0003] The method for processing safety rope monitoring data in the prior art generally has problems such as coarse signal processing granularity, weak abnormal feature extraction capability and model training insensitivity to abnormal samples. Traditional wavelet or time domain methods cannot adapt to signal characteristics, resulting in smooth processing of falling impact and low detection sensitivity.

[0004] Conventional neural networks use random initialization and uniform loss function, ignoring the distribution difference of samples in the feature space, causing the model to learn towards the normal state under sample imbalance, ultimately affecting the overall recognition accuracy and abnormal response time. SUMMARY

[0005] The present application aims to provide a safety rope operation state monitoring method and system based on double-scale differential entropy features, which solves the problems of coarse signal processing, lagging abnormal identification and high false alarm rate of existing monitoring methods.

[0006] The present application is implemented by the following technical solutions. The safety rope operation state monitoring method based on double-scale differential entropy features comprises the following steps:

[0007] S1: Collecting original stress time series data and labeling to obtain a stress data set with time series labels;

[0008] S2: Extract high-frequency impact components of the original stress time series data with time sequence labels using dynamic narrow window Gabor transform, calculate low-frequency entropy values using fixed wide window integration combined with differential entropy operators, splice the high-frequency impact components and low-frequency entropy values to construct a dual-scale differential entropy feature;

[0009] S3: Construct a neural network model and train it using the dual-scale differential entropy feature;

[0010] S4: Deploy the trained neural network model on a safety rope monitoring terminal for real-time monitoring of high-altitude work safety ropes.

[0011] Further preferably, in step S1, the original stress time series data of the safety rope is collected in real time by a strain sensor array, the strain sensors are distributed in an equidistant manner in the key load-bearing section of the safety rope, and the data collection covers all scenarios of high-altitude work. During the collection process, work videos are recorded synchronously as true value references. The original stress time series data is segmented and labeled with state categories based on the changes in personnel actions and safety rope forms in the videos. The state categories are divided into three categories: normal state, warning state, and falling state. During labeling, the original stress time series data is sliced using a sliding time window, and each slice is assigned a state category label based on the video verification results.

[0012] Further preferably, in step S2, the high-frequency impact component calculation process is as follows: based on the original stress time series data, a Gaussian window function with a window width that shrinks adaptively according to the local stress variance is used for transformation to obtain the high-frequency impact component. During the transformation process, the window width adjustment factor is dynamically calculated according to the amplitude of the stress gradient.

[0013] The low-frequency entropy value calculation process is as follows: the original stress time series data is integrated using a fixed wide window, and then the integral result is calculated using a differential entropy operator to obtain the low-frequency entropy value.

[0014] Further preferably, in step S3, the process of constructing the neural network model is as follows:

[0015] S301: Define the structure of the neural network model, which adopts a hierarchical time series feature processing architecture and contains four serial modules:

[0016] Feature input layer: receives the dual-scale differential entropy feature and maintains the time series length through zero padding;

[0017] Convolutional feature extraction layer: deploy two parallel deep separable convolution modules, each containing 3 convolution kernels with a size of 1x3. The output features are activated by ReLU and then perform maximum pooling with a size of 1x2.

[0018] The gating time sequence modeling layer: the output of the convolution feature extraction layer is input into the improved gating convolution-long short-term memory unit by time step, and the hidden state dimension is set to 64;

[0019] The attention aggregation and classification layer: the hidden state sequence output by the gating convolution-long short-term memory unit is input into the gating time attention module, a weighted aggregated state vector is generated, then a fully connected layer with an output dimension of 3 is connected, and finally three-class state probabilities are output by Softmax normalization;

[0020] S302: optimizing the initial weight of the convolution kernel in the neural network model based on the density peak value:

[0021] S303: improving the gating convolution-long short-term memory unit, embedding a lightweight convolution operation in the forgetting gate of the long short-term memory network;

[0022] S304: calculating the state probability based on attention weighting;

[0023] S305: calculating the focus loss function based on the intra-class variance;

[0024] S306: iterative updating and saving the neural network model.

[0025] Further preferably, the specific calculation process of the improved gating convolution-long short-term memory unit in step S303 is as follows:

[0026] Input gate calculation: the hidden state vector at the previous time and the input feature vector at the current time are spliced, then linearly transformed by a weight matrix and biased, and finally the input gate output value is obtained by applying the Sigmoid activation function;

[0027] Improved forgetting gate calculation: one-dimensional convolution operation is embedded in the standard forgetting gate calculation to realize the gating regulation of local feature perception. Specifically, the original fully connected term of the forgetting gate is added to the new convolution term, and then the forgetting gate output value is obtained by applying the Sigmoid activation function;

[0028] Candidate memory cell calculation: the hidden state vector at the previous time and the input feature vector at the current time are spliced, then linearly transformed by a weight matrix and biased, and finally the candidate memory cell value is obtained by applying the hyperbolic tangent activation function;

[0029] Memory update calculation: the memory cell state at the previous time controlled by the forgetting gate output value is added to the current candidate memory cell value controlled by the input gate output value, realizing the smooth migration and update of the memory cell state from the old state to the new information containing state;

[0030] Output gate calculation: the hidden state vector at the previous moment is spliced with the current input feature vector, and the output gate value is obtained through linear transformation of the weight matrix, bias and application of the Sigmoid activation function, then the output gate value is multiplied with the current memory cell state processed by the hyperbolic tangent activation function to obtain the hidden state vector at the current moment, control the output proportion of the memory cell state, and generate the cell output for the downstream task.

[0031] Further preferably, the specific process of calculating the state probability based on attention weighting in step S304 is as follows:

[0032] Calculate the importance score of the hidden state at each time step: splice the hidden state vector at the current time step with the hidden state vector at the terminal time step, and then perform dot product operation with the score vector after linear transformation of the weight matrix, bias and application of the hyperbolic tangent activation function, to obtain the importance score, wherein the score vector is used to map high-dimensional features into scalar scores;

[0033] Convert the importance score into normalized attention weight and combine with the gating mechanism: the importance score is normalized by Softmax to obtain the initial weight, and then multiplied by the gating value generated based on the current hidden state through the Sigmoid activation function to obtain the final attention weight;

[0034] Weighted aggregation of state at each time step and calculation of state category probability: the final attention weight is used to weighted sum the hidden state vectors at each time step to obtain the aggregated state vector, then the aggregated state vector is subjected to feature selection and linear transformation of each state category, and finally the Softmax function is applied to output the probability value of each state category, and the probability distribution of each state category to which the stress time series data belongs is output.

[0035] Further preferably, the specific process of calculating the focus loss function based on intra-class variance in step S305 is as follows:

[0036] Calculate the intra-class variance of each stress time series data in its real state category as the distribution dispersion;

[0037] Construct an improved focus loss function containing intra-class variance weight.

[0038] Further preferably, the specific process of real-time monitoring of high-altitude operation safety rope in step S4 is as follows:

[0039] Data preprocessing: the original stress time series data collected by the strain sensor is divided by sliding window, and the double-scale differential entropy features are obtained by executing steps S1-S2 window by window (no labeling is required during real-time monitoring process);

[0040] Real-time inference: input the trained neural network model with double-scale differential entropy features, sequentially through convolution feature extraction, gated LSTM time series modeling, and attention weighted aggregation, output the three-class state label probability of the current time window;

[0041] State decision: adopt threshold decision mechanism, fall state probability>0.8 immediately alarm, warning state probability>0.6 trigger pre-alarm, combined with time continuity constraint to suppress false alarm;

[0042] Early warning linkage: fall state alarm triggers safety rope self-locking device emergency stop protection, at the same time sends positioning information to the supervision platform through the LoRa wireless module, continuously records the stress double-scale differential entropy features and state log for periodic incremental optimization of the neural network model.

[0043] The application also provides a safety rope operation state monitoring system based on double-scale differential entropy features, which is used to realize the safety rope operation state monitoring method based on double-scale differential entropy features, and comprises a strain sensor array, a data acquisition module, a data labeling module, a feature extraction module, a model training module, a real-time monitoring module and an early warning linkage module.

[0044] Strain sensor array: distributed in the key force bearing section of the safety rope at equal intervals, used for real-time acquisition of original stress time series data of the safety rope;

[0045] Data acquisition module: connected with the strain sensor array, used for receiving the original stress time series data and synchronously acquiring the operation video as a true value reference;

[0046] Data labeling module: used for slicing the original stress time series data by using a sliding time window and labeling the state category according to the personnel action and safety rope shape change in the operation video, to obtain a stress data set with time series labeling;

[0047] Feature extraction module: used for extracting high-frequency impact components of the original stress time series data with time series labeling by using dynamic narrow window Gabor transform, calculating low-frequency entropy values by using fixed wide window integration combined with a differential entropy operator, splicing the high-frequency impact components and the low-frequency entropy values, and constructing double-scale differential entropy features;

[0048] Model training module: used for constructing a neural network model, training the neural network model through the double-scale differential entropy features, and saving the trained neural network model;

[0049] Real-time monitoring module: deployed in the safety rope monitoring terminal and connected with the strain sensor array, used for applying the trained neural network model to the real-time acquired original stress time series data, outputting the state label probability, and making a state decision;

[0050] Early warning linkage module: connected with the real-time monitoring module, used for triggering the safety rope self-locking device emergency stop protection when the falling state is monitored, and sending the positioning information to the supervision platform through the LoRa wireless module.

[0051] The application also provides an electronic device comprising a memory and a processor, wherein the memory stores computer executable instructions, and the processor invokes the computer executable instructions to perform each step of the safety rope operation state monitoring method based on double-scale differential entropy features.

[0052] The application has the following characteristics:

[0053] The application adopts a double-scale differential entropy feature extraction mechanism, dynamically adjusts the Gabor transform window width, adaptively captures high-frequency sudden impact and low-frequency periodic fluctuation signals, and effectively distinguishes the normal, warning and falling three states;

[0054] In the neural network model training, the initial weights of the convolution kernel in the neural network model are optimized based on the density peak value, so that the neural network model pays attention to the typical feature mode of the abnormal sample from the early training stage, and the recognition ability of the neural network model for sparse falling data is significantly improved;

[0055] The improved gated convolution-long short-term memory unit is fused in the neural network model, the lightweight convolution operation is embedded in the long short-term memory network (LSTM) forgetting gate, so that the neural network model has the ability to process short-time mutation features and model long-time state dependence, and the understanding of the high-altitude falling state evolution process is improved;

[0056] The gated attention is adopted, so that the neural network model can automatically focus on the time period with key risk features, and the problem that the traditional model is not sensitive to important moments is relieved;

[0057] The focal loss function is improved based on the within-class variance weight mechanism, the focal loss is combined to model the high-variance samples in the abnormal state, and the learning ability and generalization level of the neural network model for complex abnormal patterns are improved.

[0058] Based on the above improvements, the application significantly improves the accuracy and robustness of the high-altitude operation safety rope state recognition, and the reasoning speed is more than five times faster than that of the traditional model, which meets the real-time monitoring demand; in the state decision, the system suppresses false positives through the multi-window continuity constraint, and realizes the fast, stable and reliable falling early warning linkage mechanism combined with the high-precision capture of stress mutation. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The application provides a safety rope operation state monitoring method based on double-scale differential entropy features.

[0060] Figure 2Traditional wavelet feature distribution map.

[0061] Figure 3 Dual-scale differential entropy feature distribution map.

[0062] Figure 4 Performance comparison chart of different methods in various states.

[0063] Figure 5 Real-time processing performance comparison chart of different methods. DETAILED DESCRIPTION

[0064] The application will be further illustrated in detail below with reference to the examples.

[0065] Example 1

[0066] Reference Figure 1 The safety rope operation state monitoring method based on the dual-scale differential entropy feature has the following steps:

[0067] S1: Collecting original stress time series data and labeling to obtain a stress data set with time series labels;

[0068] The original stress time series data of the safety rope is collected in real time by a high-precision strain sensor array. The high-precision strain sensors are distributed in an equal-interval manner in the key load-bearing sections of the safety rope, and the sampling frequency is set to 1 kHz to ensure the capture of microsecond-level impact signals during the falling moment. The data collection covers all scenarios of aerial work, including typical working conditions such as normal climbing, tool operation, and sudden falling. A work video is recorded synchronously during the collection process as a true value reference. The original stress time series data is segmented and labeled according to the state category based on the changes in personnel actions and safety rope shapes in the video. The state category is divided into three categories: normal state (corresponding to low-amplitude periodic fluctuations during stable climbing or static suspension), warning state (corresponding to medium-amplitude transients caused by rapid displacement or instrument collision), and falling state (corresponding to high-amplitude peak impact during sudden tightening of the safety rope). During labeling, the original stress time series data is sliced using a sliding time window (window length 0.5 seconds, step length 0.1 seconds). Each slice is assigned a state category label according to the video verification result, and finally a stress data set with time series labels is formed.

[0069] S2: Extracting high-frequency impact components of the original stress time series data with time series labels using dynamic narrow-window Gabor transform, calculating low-frequency entropy values using fixed wide-window integration combined with differential entropy operators, concatenating the high-frequency impact components and low-frequency entropy values, and constructing dual-scale differential entropy features;

[0070] S3: Constructing a neural network model and training it using dual-scale differential entropy features;

[0071] S4: The trained neural network model is deployed on the safety rope monitoring terminal for real-time monitoring of the safety rope of aerial work, and the process is as follows:

[0072] Data preprocessing: The original stress time series data collected by the strain sensor is divided by a sliding window (window length 2 seconds, step length 0.1 seconds), and the double-scale differential entropy features are obtained by executing steps S1-S2 window by window;

[0073] Real-time inference: The double-scale differential entropy features are input into the trained neural network model, and the three-class state label probabilities of the current time window are output by sequentially performing convolution feature extraction, gated LSTM time series modeling and attention weighted aggregation;

[0074] State decision: A threshold decision mechanism (falling state probability > 0.8 to immediately alarm, warning state probability > 0.6 to trigger a pre-alarm) is adopted, and a time continuity constraint (continuous 3-window pre-alarm takes effect) is used to suppress false alarms;

[0075] Early warning linkage: The falling state alarm triggers the safety rope self-locking device to stop protection, and at the same time sends the positioning information to the supervision platform through the LoRa wireless module. The double-scale differential entropy features and state logs of the stress are continuously recorded for periodic incremental optimization of the neural network model.

[0076] The safety rope stress data contains both high-frequency sudden impact at the moment of falling and low-frequency periodic fluctuations during normal operation, making it difficult for conventional time-domain statistical features to effectively distinguish between sudden abnormal events and noise generated by normal actions. The traditional wavelet transform method needs to pre-set a fixed basis function and cannot adaptively adjust the time-frequency resolution of high-frequency sudden signals according to the signal characteristics, resulting in the key falling peak feature being smoothed during analysis, reducing the sensitivity of abnormal detection. The present application constructs a double-scale differential entropy feature with an adaptive window length, which separates the impact and periodic components by dynamically adjusting the window function width of Gabor transform, and the specific steps are as follows:

[0077] S2-1: Extract the high-frequency impact component of the original stress time series data using dynamic narrow window Gabor transform. Specifically, based on the original stress time series data, a Gaussian window function with a window width that automatically shrinks with the local stress variance is used for transformation, which can focus on high-frequency sudden signals and obtain a high-frequency impact component. The window width is automatically narrowed in the region with rapid signal change to improve the time resolution, which is expressed as:

[0078] ;

[0079] In the formula, is the high-frequency impact component; is the original stress time series data; is the time index; is the time index different from t, and j is the imaginary unit, is the angular frequency, is the Gaussian window function under the dynamic short window width condition, and the calculation method is represented as ; is the dynamic short window width, and the calculation method is represented as ; is the variance of the stress within the time window ; is the local time window length, such as = 0.1 seconds; is the window width adjustment factor; is the stress time series data within the time window ;

[0080] The window width adjustment factor is dynamically calculated according to the amplitude of the stress gradient , specifically, the stress gradient is calculated using the difference between the current stress and the previous stress, and the window width adjustment factor is obtained by mapping based on the amplitude of the stress gradient through the Sigmoid function, so that the window width is automatically reduced in the stress mutation area, thereby improving the time resolution to accurately capture the impact signal, represented as:

[0081] ;

[0082] In the formula, is the stress gradient at time t, and the calculation method is represented as , is the stress value at time t, is the stress value at time ; is the window width adjustment steepness coefficient, such as ; is the L2 norm, used to represent the amplitude of the stress gradient.

[0083] S2-2: Fixed wide window integration combined with differential entropy operator is used to calculate the low-frequency entropy value, specifically, the original stress time series data is integrated using a fixed wide window, and then the integral result is calculated using a differential entropy operator, which quantifies the low-frequency periodic fluctuation component while avoiding the problem of spectral leakage, and obtains the low-frequency entropy value, represented as:

[0084] ;

[0085] In the formula, is the low-frequency entropy value, is the Gaussian window function under the fixed long window width condition; is the fixed long window width, and the calculation method is represented as ; is the differential entropy operator, and the calculation method is represented as ; To fix the result of the wide window integration; To log function, the default base is natural constant; To numerical stability term, such as, , to prevent zero value log calculation overflow.

[0086] S2-3: splice the high-frequency impact component and the low-frequency entropy value to construct a dual-scale differential entropy feature containing dual-scale information, denoted as:

[0087] ;

[0088] In the formula, is a feature splicing operation; is a dual-scale differential entropy feature.

[0089] In one embodiment, feature space distribution comparison analysis is performed to verify the effectiveness of the dual-scale differential entropy feature extraction method proposed in the application. By comparing the distribution of the traditional wavelet feature and the joint feature tensor in the two-dimensional space, the difference in feature separability is intuitively displayed, Figure 2 The traditional wavelet feature space shows that the normal state (blue points), the warning state (orange points) and the falling state (green points) samples are highly overlapped, especially in the feature dimension 1 value range 0.3-0.6 and the feature dimension 2 value range 0.2-0.5 region. There is serious aliasing, which shows that the traditional method is difficult to distinguish the key features of different states, Figure 3 The dual-scale differential entropy feature distribution of the application is shown. The normal state is densely gathered in the lower left area (feature dimension 1≈0.2, feature dimension 2≈0.3), the warning state is linearly distributed in the upper right area (feature dimension 1≈0.5, feature dimension 2≈0.7), and the falling state is independently concentrated in the lower right area (feature dimension 1≈0.8, feature dimension 2≈0.4). The three states form a clear separated clustering cluster. The experimental results prove that the application separates the impact and periodic components through the dynamic window width mechanism (the window width is adaptively shrunk according to the local stress variance), so that the feature space class distance is significantly expanded and the class aggregation degree is obviously improved, which lays a foundation for subsequent accurate classification.

[0090] In step S3 of the embodiment, the process of constructing the neural network model is as follows:

[0091] S301: define the structure of the neural network model;

[0092] The neural network model adopts a hierarchical time series feature processing architecture, which includes four serial modules:

[0093] Feature input layer: receives the dual-scale differential entropy feature (dimension: time step x feature channel) and keeps the time series length unchanged through zero padding;

[0094] Convolutional feature extraction layer: two sets of parallel deep separable convolution modules are deployed, each set of deep separable convolution modules includes 3 convolution kernels (size 1x3), which respectively focus on local pattern mining of high-frequency impact components and low-frequency entropy values, and the output features are activated by ReLU and then perform maximum pooling (pooling kernel size 1x2) to compress redundant information;

[0095] Gated recurrent modeling layer: the output of the convolutional feature extraction layer is input into the improved gated convolution-long short-term memory unit (hidden state dimension 64) according to the time step, the gated recurrent modeling layer captures short-time mutation features through the forgetting gate convolution mechanism, and maintains long-time state dependency relationship by using the memory unit;

[0096] Attention aggregation and classification layer: the hidden state sequence output by the gated convolution-long short-term memory unit is input into the gated recurrent attention module, a weighted aggregated state vector is generated, then a fully connected layer (output dimension 3) is connected, and finally the three state probabilities of normal state, warning state and falling state are output through Softmax normalization.

[0097] S302: optimizing the initial weight of the convolution kernel in the neural network model based on the density peak value;

[0098] Each stress time series data is regarded as a sample, the data samples of the normal operation state of the safety rope account for a high proportion in the training set, and the abnormal samples are sparse and diverse in form. The convolution kernel initialized at random is prone to ignore these rare risk patterns, so that the neural network model is not sensitive to abnormal features at the beginning of training. The traditional initialization method assumes that the data obeys Gaussian distribution, lacks the ability to model the clustering center of the multi-modal stress features of the safety rope, and it is difficult to effectively focus the initial weight on the typical abnormal pattern. The present application determines the prior distribution of the convolution kernel by using the density peak value of the feature space, so that the initial weight focuses on the typical risk pattern, and the specific steps are as follows:

[0099] S302-1, density peak detection: the clustering center in the feature space is identified by calculating the local density of each sample and the minimum distance of each sample to the higher density sample. Specifically, based on the two-scale differential entropy features of all samples, the local density and minimum distance of each sample are calculated, and then the samples meeting the conditions are screened out as the clustering center according to the set density threshold and distance threshold. The density peak point representing the typical mode in the feature space is identified as the initial clustering center, which is represented as:

[0100] ;

[0101] ;

[0102] In the formula, is the two-scale differential entropy feature of the i th sample; Let be the dual-scale differential entropy feature of the j-th sample; The total number of samples in the training set; The bandwidth is calculated based on density, and the calculation method is expressed as follows: ; This is the median operator, used to estimate the distance to typical features; Let be the local density of the i-th sample; a larger value indicates that the surrounding samples are more densely packed. It represents the minimum distance from the i-th sample to the higher-density sample;

[0103] Density threshold This indicates quantile operations. This represents the local density of each sample, taken as the 90th quantile of the density distribution; As the distance threshold, the 90th percentile of the distance distribution is used. This represents the minimum distance from each sample to higher-density samples. Further, the selection criterion is: when... and When the i-th sample is selected as the cluster center ,definition Cluster center samples selected for density peak detection.

[0104] S302-2, Neighborhood Sample Aggregation: Using the identified cluster centers as base points, samples within a certain range around them are collected to form neighborhoods. Specifically, the characteristic distance from each sample to the cluster center is calculated, and samples whose distance is less than or equal to the density calculation bandwidth are assigned to the neighborhood of that cluster center. A set of samples within the local neighborhood of each cluster center is determined, represented as:

[0105] ;

[0106] In the formula, For the first The neighborhood of each cluster center; For sample index; This is a condition symbol, indicating that the condition is met.

[0107] S302-3, Initial weight calculation of convolution kernels: Initial weights of convolution kernels are generated based on the feature mean of neighborhood samples and with the addition of a controllable perturbation. Specifically, the mean vector of features of all samples in the neighborhood is calculated, the mean is subtracted from the features of each sample in the neighborhood, and then the average of these subtracted features is calculated. Finally, a uniformly distributed random perturbation is superimposed to generate initial weights for each convolution kernel that focus on the feature patterns of its corresponding cluster center neighborhood, expressed as:

[0108] ;

[0109] In the formula, initial weights of the first convolution kernel; number of samples within the neighborhood mean vector of features within the neighborhood , which is calculated as uniform distribution random disturbance, range parameter of the uniform distribution, such as disturbance coefficient, such as , controls the strength of the random disturbance. It should be noted that represents the mean subtraction operation, which is used to eliminate the within-class common components and retain the discriminative features.

[0110] S303, improved gated convolution-long short-term memory unit;

[0111] The aerial work safety rope state classification task needs to capture the instantaneous stress response and subsequent continuous state change in the falling event, and requires the model to have the ability to process local mutation features and long-range state dependence. The gating mechanism of the standard long short-term memory network (LSTM) ignores the correlation between local features of input data, resulting in a lag in its response to sudden impact signals. Although CNN is good at extracting local features, it is difficult to effectively model the long-time scale state transition process, and cannot fully process the inherent time sequence dependence of the safety rope state conversion. The present application embeds a lightweight convolution operation in the forget gate of the long short-term memory network (LSTM), realizes dynamic gating adjustment of local feature perception, and thus synchronously captures short-time mutations and long-time dependencies. The response to recent local patterns is strengthened through the convolution term, and the forgetting gate strength is dynamically adjusted to avoid oscillation noise interference in the falling judgment. The specific steps are as follows:

[0112] 1) Input gate calculation

[0113] The input gate value is calculated to control the degree of new information flowing into the memory unit. Specifically, the hidden state vector at the previous moment is spliced with the input feature vector at the current moment, and after linear transformation through a weight matrix and bias, the input gate output value is obtained by applying the Sigmoid activation function. A weight scalar is generated to determine the contribution size of the current input information to updating the memory unit, which is represented as:

[0114] ;

[0115] In the formula, represents the input gate output value at the moment, which is a scalar with a dimension of , and the value range is , which is used to control the inflow amount of new information;​​​​ This represents the Sigmoid activation function, which maps the input to... interval; The weight matrix of the input gate has dimensions of . ,in For the hidden state dimension, such as, , The number of convolution channels, for example, , used for linear transformations; This represents the vector concatenation operation, which... and splicing into dimensions ; express The hidden state vector at time t, with dimension . It carries historical timeline information; express The input feature vector at time step 1, i.e., the two-scale differential entropy feature. The slice at time t is specifically as follows: , where the high-frequency impact component and the low-frequency entropy value each occupy one dimension; This represents the bias vector of the input gate, used to adjust the activation threshold.

[0116] 2) Improved calculation of forget gate

[0117] A one-dimensional convolution operation is embedded in the standard forget gate computation to achieve local feature awareness gating. Specifically, the original fully connected terms of the forget gate are added to the newly added convolutional terms, and then the output value of the forget gate is obtained through the sigmoid activation function. This allows the forget gate to dynamically perceive recent local feature patterns and reduce its output value when high-frequency spikes are detected, thereby strengthening the forgetting of old states and avoiding noise interference. This can be expressed as:

[0118] ;

[0119] In the formula, express The output value of the forget gate at each time step is a scalar with a range of values. It is used to control the degree to which historical memory is preserved; The weight matrix represents the forget gate, used for the linear transformation of the fully connected part; This represents the learnable weight vector of the convolution term, used to scalarize the convolution output; This represents a matrix multiplication operation, where the convolution output is multiplied by... Multiplying them together yields a scalar value; This represents a one-dimensional convolution operation with a kernel width of . Input is a time window The output dimension is , used to extract recent local features; denotes the input feature sequence from to time, with dimension , where is the convolution kernel width, such as ; denotes the bias vector of the forget gate. It is worth noting that term is specifically designed to handle recent step data, output high value when detecting high-frequency peaks such as fall initiation, strengthen forgetting old state, reduce , avoid oscillation noise interference.

[0120] 3) Candidate memory cell calculation

[0121] Generate candidate memory cell values to carry new information of current input. Specifically, concatenate the hidden state vector of the previous time with the input feature vector of the current time, apply the hyperbolic tangent activation function after linear transformation by weight matrix and bias, to obtain the candidate memory cell value, calculate and normalize the possible new memory content at the current time, denoted as:

[0122] ;

[0123] In the formula, denotes the candidate memory cell value at time t, which carries the new information of the current input; denotes the hyperbolic tangent function; denotes the bias vector of the candidate memory.

[0124] 4) Memory update calculation

[0125] Update the memory cell state by combining the forget gate and the input gate. Specifically, add the memory cell state of the previous time controlled by the output value of the forget gate, and the current candidate memory cell value controlled by the output value of the input gate, to realize the smooth transition and update of the memory cell state from the old state to the new information containing state, denoted as:

[0126] ;

[0127] In the formula, denotes the updated memory cell state at time t, which integrates historical and current information; denotes the memory cell state at time, which carries the accumulated historical information; denotes the element-wise multiplication operation.

[0128] 5) Output gate calculation

[0129] The output gate value and the current hidden state vector are calculated, specifically, the hidden state vector at the previous moment and the current input feature vector are spliced, the output gate value is obtained through linear transformation of a weight matrix, bias and application of a Sigmoid activation function, then the output gate value is multiplied by the current memory cell state processed by a hyperbolic tangent activation function to obtain the hidden state vector at the current moment, the proportion of the output of the memory cell state is controlled, and a cell output for a downstream task is generated, represented as:

[0130] ;

[0131] ;

[0132] In the formula, the output gate output value at t moment, dimension is a scalar, value range , controlling the proportion of the output of the memory cell; the hidden state vector at t moment, as a cell output for a downstream task; the weight matrix of the output gate, used for linear transformation; the bias vector of the output gate.

[0133] S304, state probability calculation based on attention weighting;

[0134] The safety rope state conversion has a time sequence dependency relationship, the standard classification layer usually directly compresses the hidden state at the end moment, and the information of the key intermediate state in the sequence is lost, the conventional global pooling operation cannot highlight the discriminative state fragment in the risk accumulation stage, resulting that the model has insufficient feature response to the key period of the falling early warning, and the timeliness and accuracy of the early warning are affected. The application adopts the gated attention, dynamically weights and aggregates the states at each time step, and combines state category exclusive feature selection to realize state-sensitive probability calculation, learns the state importance and credibility jointly, automatically improves the weight in the risk accumulation period, and makes different state categories pay attention to different feature patterns through a feature selection vector. The specific steps are as follows:

[0135] 1) State importance score

[0136] The importance score of the hidden state at each time step is calculated, specifically, the hidden state vector at the current time step and the hidden state vector at the end time step are spliced, linear transformation of a weight matrix, bias and application of a hyperbolic tangent activation function are performed, and then dot product operation is performed with a score vector to obtain a scalar score, the relative importance of each time step state to the final classification decision is evaluated, represented as:

[0137] ;

[0138] In the formula, denotes the state importance score at time t, the larger the value, the more critical the state at this time to classification; denotes the score vector, which is a trainable parameter, used to map high-dimensional features to a scalar score; is the transpose of ; denotes the weight matrix, used for feature transformation; is the terminal time hidden state, carrying global timing information; denotes the vector concatenation operator; denotes the bias vector.

[0139] 2) Gating attention weight calculation

[0140] Convert the importance score into normalized attention weights and combine the gating mechanism. Specifically, the importance score is normalized by Softmax to obtain the initial weight, which is then multiplied by the gating value generated by the Sigmoid activation function based on the current hidden state to obtain the final attention weight, which reflects the importance of the state and can suppress the influence of noise period, denoted as:

[0141] ;

[0142] In the formula, denotes the final attention weight at time t, with a value range of , used for state aggregation; denotes the natural exponential function; is the state importance score at time step ; denotes the exponential sum of all time step scores, which realizes Softmax normalization; denotes the gating weight vector, used to generate state credibility scores; denotes the gating bias.

[0143] 3) State aggregation and probability calculation

[0144] Weighted aggregation of states at each time step and calculation of state category probability. Specifically, the final attention weight is used to weight and sum the hidden state vectors at each time step to obtain an aggregated state vector. Then, feature selection and linear transformation are performed on the vector for each state category. Finally, the Softmax function is applied to output the probability value of each state category, and the probability distribution of the sample belonging to each state category is output, denoted as:

[0145] ;

[0146] ;

[0147] In the formula, to aggregate the weighted state vector, and integrate the key information of each period; is the total length of the time series; is the probability distribution vector of the cth state category, with a dimension of is a three-dimensional vector with a value range of 0 to 1, corresponding to the normal state, the warning state, and the falling state; represents the Softmax function; represents the state category weight vector, which is used for linear transformation; represents the state category-specific feature selection vector, which realizes class-conditioned feature enhancement; represents the state category bias.

[0148] S305, calculate the focus loss function based on the intra-class variance;

[0149] In the safety rope monitoring data, the tension of the normal state sample is stable, and the feature distribution is usually concentrated, while the feature distribution of the abnormal sample has a large variance, the falling form is various, and the proportion in the total sample is extremely low. The standard focus loss function mainly adjusts the inter-class weight to solve the problem of unbalanced sample quantity, but does not consider the distribution difference of the same class sample, which leads to insufficient learning of the model to the feature distribution of the abnormal pattern with high dispersion and various forms in the training process, affecting the generalization ability of the abnormal detection. The focus loss function is improved based on the intra-class variance weight mechanism in the present application, and the attention degree to the high-variance abnormal sample is increased, while the inter-class imbalance processing ability is retained. The specific steps are as follows:

[0150] 1) Intra-class variance calculation

[0151] The intra-class variance of each sample in its true state category is calculated as the distribution dispersion, specifically, for the current sample, the average value of the square distance between the weighted aggregation state vector of the sample and the weighted aggregation state vector of all other samples in the state category to which the sample belongs is calculated, which quantifies the dispersion degree of the current sample feature relative to the center of the same class sample, and is represented as:

[0152]

[0153] In the formula, is the intra-class variance of the ith sample, representing the distribution dispersion of the same class sample; is the sample set of the state category to which the ith sample belongs; is the total number of samples of the state category to which the ith sample belongs; is the weighted aggregation state vector of the ith sample; is the weighted aggregation state vector of the jth sample.

[0154] 2) Loss function construction

[0155] ​An improved focal loss function is constructed, which contains the intra-class variance weight. Specifically, on the basis of the calculation formula of the standard focal loss, the intra-class variance is embedded as an exponential term, which dynamically adjusts the loss contribution of each sample. The loss function not only pays attention to the difficulty of sample classification and the class imbalance problem, but also especially increases the attention to high-variance abnormal samples, promotes the model to learn diversified abnormal patterns, and is expressed as:

[0156] ;

[0157] In the formula, is an improved focal loss function; is the probability value of the i-th sample in the real state class ; is a hierarchical inter-class weight coefficient, which is configured according to the risk level, such as ; is a difficult sample focusing parameter, such as , which increases the punishment of misclassified samples; is an intra-class variance adjustment coefficient, such as , which controls the strength of variance influence.

[0158] S306, neural network model iterative update and neural network model saving;

[0159] The neural network model training adopts small batch random gradient descent (batch size 256), the initial learning rate is set to 0.01, and the cosine annealing strategy is applied for dynamic decay.

[0160] After each round of training, the focal loss function is calculated using an independent validation set. When the validation loss does not decrease for 10 consecutive rounds, the early stopping mechanism is triggered and the optimal weight is rolled back.

[0161] The neural network model saving is divided into two versions: the final model (parameter snapshot at the end of training) and the optimal validation model (parameters when the validation loss is the lowest), both of which are stored in the network structure and weight in HDF5 format.

[0162] In one embodiment, state detection performance comparison analysis is performed to evaluate the detection performance of different methods on various states, such as Figure 4As shown, the abscissa represents the state category (normal / warning / fall), and the ordinate shows the F1 score (comprehensive evaluation index, value range 0-1). The bar chart clearly presents the comparison of the five methods. The traditional support vector machine (SVM) performs the worst (the lowest column) in fall state detection. The random forest is slightly improved. The long short-term memory network (LSTM) and the convolution-long short-term memory fusion model (CNN-LSTM) are improved in turn. The method of the present application (the rightmost column group) has a significant advantage in the three states, especially in the fall state detection, and the column height of the method of the present application is much higher than that of other methods (close to the position of the ordinate 1.0). The warning state detection also presents similar advantages. It is worth noting that the performance of the traditional method in the warning and fall states is greatly different (the height difference of the adjacent columns is obvious). The column heights of the three states of the method of the present application tend to be consistent, indicating that the convolution kernel initialization based on the density peak value makes the model focus on the typical risk pattern, and the gated attention effectively captures the risk accumulation features, solving the detection blind area problem caused by the sparsity of abnormal samples.

[0163] In one embodiment, real-time processing performance comparison and analysis of different methods are carried out, as shown in FIG. 6. Figure 5 As shown, a double ordinate design is adopted. The left ordinate represents the processing time (unit: millisecond), and the right ordinate represents the accuracy (value range 0.8-1.0). The five groups of bar charts show that the traditional support vector machine (SVM) has a medium processing time but the lowest accuracy (the first group on the left), the random forest has a shorter processing time but insufficient accuracy, the long short-term memory network (LSTM) has improved accuracy but significantly increased processing time (the highest column), the convolution-long short-term memory fusion model (CNN-LSTM) is a compromise between processing time and accuracy, and the method of the present application presents a breakthrough advantage, with the lowest processing time column and the highest accuracy broken line vertex. The processing time of the method of the present application is only one fifth of that of the long short-term memory network, but the accuracy is improved by more than five percentage points, indicating that the dynamic window width calculation reduces redundant frequency spectrum analysis, the lightweight gated convolution reduces the unit calculation complexity, and the feature space pre-aggregation optimizes the attention weight calculation efficiency, indicating that the method of the present application meets the dual strict requirements of real-time and reliability in safety monitoring scenarios.

[0164] Embodiment 2

[0165] The embodiment provides a safety rope operation state monitoring system based on a double-scale differential entropy feature, which is used to implement the safety rope operation state monitoring method based on the double-scale differential entropy feature in embodiment 1, and comprises a strain sensor array, a data acquisition module, a data labeling module, a feature extraction module, a model training module, a real-time monitoring module and a pre-warning linkage module.

[0166] The strain sensor array is distributed in an equal-interval manner in a key load-bearing section of the safety rope and is used to acquire original stress time series data of the safety rope in real time.

[0167] Data acquisition module: connected with the strain sensor array, used to receive the original stress time series data, and simultaneously collect the operation video as the true value reference;

[0168] Data labeling module: used to slice the original stress time series data and label the state category by using a sliding time window according to the personnel action and safety rope morphological change in the operation video, to obtain the stress data set with time series labeling;

[0169] Feature extraction module: used to extract the high-frequency impact component of the original stress time series data with time series labeling by using dynamic narrow window Gabor transform, and calculate the low-frequency entropy value by using fixed wide window integral combined with differential entropy operator, to splice the high-frequency impact component and the low-frequency entropy value, and construct the double-scale differential entropy feature;

[0170] Model training module: used to construct a neural network model, train the neural network model by using the double-scale differential entropy feature, and save the trained neural network model;

[0171] Real-time monitoring module: deployed in the safety rope monitoring terminal, connected with the strain sensor array, used to apply the trained neural network model to the real-time collected original stress time series data, output the state label probability, and make state decision;

[0172] Early warning linkage module: connected with the real-time monitoring module, used to trigger the safety rope self-locking device emergency stop protection when the falling state is monitored, and send the positioning information to the supervision platform through the LoRa wireless module.

[0173] Embodiment 3

[0174] The embodiment provides an electronic device, including a memory and a processor, the memory stores computer executable instructions, and the processor invokes the computer executable instructions to execute each step of the safety rope operation state monitoring method based on the double-scale differential entropy feature described in embodiment 1.

[0175] The above only expresses the preferred embodiments of the present application, and is not other forms of the present application Limitations, any skilled in the art of technical personnel may use the above disclosed content to change or modify the equivalent embodiment of equivalent changes. However, any simple modification, equivalent change and modification of the above embodiments according to the technical essence of the present application without departing from the technical scheme content of the present application, still belongs to the protection scope of the present application technical scheme.

Claims

1. A method for monitoring the operational status of safety ropes based on dual-scale differential entropy features, characterized in that, Includes the following steps: S1: Collect raw stress time series data and annotate it to obtain a stress dataset with time series annotations; S2: The high-frequency impact component of the original stress time series data with time series labeling is extracted by dynamic narrow window Gabor transform. The low-frequency entropy value is calculated by using fixed wide window integral combined with differential entropy operator. The high-frequency impact component and the low-frequency entropy value are concatenated to construct a dual-scale differential entropy feature. The calculation process of the high-frequency impact component is as follows: based on the original stress time series data, a Gaussian window function with adaptive shrinkage of window width as local stress variance is used to transform the high-frequency impact component. During the transformation process, the window width adjustment factor is dynamically calculated according to the magnitude of the stress gradient. The low-frequency entropy value is calculated as follows: Gaussian window integration is performed on the original stress time series data using a fixed wide window, and then the differential entropy operator is applied to the integration result to calculate the low-frequency entropy value. S3: Construct a neural network model and train it using dual-scale differential entropy features; S4: Deploy the trained neural network model on the safety rope monitoring terminal to perform real-time monitoring of safety ropes for high-altitude operations; The process of constructing the neural network model in step S3 is as follows: S301: Define the neural network model structure. The neural network model adopts a hierarchical temporal feature processing architecture, which includes four levels of concatenated modules: Feature input layer: Receives dual-scale differential entropy features and maintains the temporal length unchanged through zero padding; Convolutional feature extraction layer: Two sets of parallel depthwise separable convolutional modules are deployed. Each set of depthwise separable convolutional modules contains 3 convolutional kernels of size 1×3. The output features are activated by ReLU and then subjected to max pooling of size 1×2. Gated temporal modeling layer: The output of the convolutional feature extraction layer is input into the improved gated convolutional-long short-term memory unit according to time steps, and the hidden state dimension is set to 64. Attention aggregation and classification layer: The hidden state sequence output by the gated convolution-long short-term memory unit is input into the gated temporal attention module, which generates a weighted aggregated state vector and then connects it to a fully connected layer with an output dimension of 3. Finally, it is normalized by Softmax to output the probabilities of the three states. S302: Optimizing the initial weights of convolutional kernels in a neural network model based on peak density: S303: Improved gated convolution - Long Short-Term Memory unit, embedding lightweight convolution operations in the forget gate of the Long Short-Term Memory network; S304: Attention-weighted state probability calculation; S305: Calculate the focus loss function based on within-class variance; S306: Iterative updates and saving of neural network models.

2. The safety rope operation status monitoring method based on dual-scale differential entropy features according to claim 1, characterized in that, In step S1, the original stress time series data of the safety rope is collected in real time through a strain sensor array. The strain sensors are distributed at equal intervals in the key load-bearing sections of the safety rope. The data collection covers the entire high-altitude operation scenario. During the collection process, the operation video is recorded simultaneously as a true value reference. Based on the personnel's actions and the changes in the shape of the safety rope in the video, the original stress time series data is segmented and labeled with state categories. There are three state categories: normal state, warning state, and fall state. When labeling, a sliding time window is used to slice the original stress time series data. Each slice is assigned a state category label according to the video verification results.

3. The safety rope operation status monitoring method based on dual-scale differential entropy features according to claim 1, characterized in that, The specific calculation process of the improved gated convolution-long short-term memory unit in step S303 is as follows: Input gate computation: The hidden state vector of the previous time step is concatenated with the input feature vector of the current time step. After linear transformation and biasing of the weight matrix, the Sigmoid activation function is applied to obtain the input gate output value. Improved calculation of the forget gate: A one-dimensional convolution operation is embedded in the standard forget gate calculation to achieve local feature awareness gating adjustment. Specifically, the original fully connected terms of the forget gate are added to the newly added convolution terms, and then the output value of the forget gate is obtained through the Sigmoid activation function. Candidate memory unit calculation: The hidden state vector of the previous time step is concatenated with the input feature vector of the current time step. After linear transformation and biasing of the weight matrix, the hyperbolic tangent activation function is applied to obtain the candidate memory unit value. Memory update calculation: The previous state of the memory cell controlled by the output value of the forget gate is added to the current candidate memory cell value controlled by the output value of the input gate, so as to achieve a smooth migration and update of the memory cell state from the old state to the state containing new information; Output gate calculation: The hidden state vector of the previous time step is concatenated with the current input feature vector. The output gate value is obtained by linear transformation of the weight matrix, biasing, and applying the Sigmoid activation function. Then, this output gate value is multiplied with the current memory cell state processed by the hyperbolic tangent activation function to obtain the hidden state vector of the current time step. This controls the output ratio of the memory cell state and generates the cell output for downstream tasks.

4. The safety rope operation status monitoring method based on dual-scale differential entropy characteristics according to claim 1, characterized in that, The specific process of calculating the state probability based on attention weighting in step S304 is as follows: Calculate the importance score of the hidden state at each time step: concatenate the hidden state vector of the current time step with the hidden state vector of the last time step, perform linear transformation of the weight matrix, bias and apply the hyperbolic tangent activation function, and then perform a dot product operation with the score vector to obtain the importance score. The score vector is used to map high-dimensional features to scalar scores. The importance score is transformed into a normalized attention weight and combined with a gating mechanism: the importance score is normalized by Softmax to obtain the initial weight, and then multiplied by the gating value generated by the Sigmoid activation function based on the current hidden state to obtain the final attention weight. Weighted aggregation of states at each time step and calculation of state category probabilities: The hidden state vectors at each time step are weighted and summed using the final attention weights to obtain an aggregated state vector. Then, feature selection and linear transformation for each state category are performed on the aggregated state vector. Finally, the Softmax function is applied to output the probability value of each state category and the probability distribution of stress time series data belonging to each state category is output.

5. The safety rope operation status monitoring method based on dual-scale differential entropy features according to claim 1, characterized in that, The specific process for calculating the focus loss function based on within-class variance in step S305 is as follows: Calculate the within-class variance of each stress time series data within its true state category as the distribution dispersion. Construct an improved focus loss function that includes in-class variance weights.

6. The safety rope operation status monitoring method based on dual-scale differential entropy characteristics according to claim 3, characterized in that, The specific procedure for real-time monitoring of the safety rope for high-altitude operations in step S4 is as follows: Data preprocessing: The raw stress time series data collected by the strain sensor is segmented by a sliding window, and steps S1-S2 are performed window by window to obtain dual-scale differential entropy features; Real-time inference: The dual-scale differential entropy features are input into the trained neural network model, which is then processed sequentially through convolutional feature extraction, gated LSTM temporal modeling, and attention-weighted aggregation to output the probabilities of the three state labels in the current time window. Status decision: A threshold decision mechanism is adopted. If the probability of a fall state is >0.8, an alarm is triggered immediately. If the probability of a warning state is >0.6, a pre-alarm is triggered. Combined with time continuity constraints, false alarms are suppressed. Early warning linkage: The fall alarm triggers the emergency stop protection of the safety rope self-locking device. At the same time, the location information is sent to the monitoring platform via the LoRa wireless module. The dual-scale differential entropy characteristics of stress and the status log are continuously recorded for periodic incremental optimization of the neural network model.

7. A safety rope operation status monitoring system based on dual-scale differential entropy characteristics, characterized in that, The method for monitoring the status of safety rope operations based on dual-scale differential entropy features as described in any one of claims 1-6 includes a strain sensor array, a data acquisition module, a data annotation module, a feature extraction module, a model training module, a real-time monitoring module, and an early warning linkage module. Strain sensor array: Distributed at equal intervals in key load-bearing sections of the safety rope to collect raw stress time-series data of the safety rope in real time; Data acquisition module: connected to the strain sensor array, used to receive raw stress time series data, and simultaneously acquire operation video as a true value reference; Data annotation module: Based on the personnel's actions and the changes in the shape of the safety rope in the operation video, the module uses a sliding time window to slice the original stress time series data and annotate the state categories to obtain a stress dataset with time series annotations. Feature extraction module: Used to extract high-frequency impact components from raw stress time series data with time series labels using dynamic narrow window Gabor transform, and to calculate low-frequency entropy values ​​using fixed wide window integral combined with differential entropy operator. The high-frequency impact components and low-frequency entropy values ​​are then concatenated to construct dual-scale differential entropy features. Model training module: used to build neural network models, train them using dual-scale differential entropy features, and save the trained neural network models; Real-time monitoring module: Deployed in the safety rope monitoring terminal, connected to the strain sensor array, it is used to apply the trained neural network model to the raw stress time series data collected in real time, output the state label probability, and make state decisions; Early warning linkage module: Connected to the real-time monitoring module, it is used to trigger the safety rope self-locking device for emergency stop protection when a fall is detected, and at the same time, it sends the location information to the monitoring platform via the LoRa wireless module.

8. An electronic device comprising a memory and a processor, the memory storing computer-executable instructions, characterized in that, The processor invokes computer-executable instructions to execute each step of the safety rope operation status monitoring method based on dual-scale differential entropy features as described in any one of claims 1-6.

Images