Intelligent monitoring method and system for dangerous state of safety belt for aerial work, and electronic device

Abstract

The present application belongs to the technical field of intelligent monitoring of safety belts, and relates to a method and system for intelligent monitoring of dangerous states of safety belts for aerial work, and electronic equipment. The present application collects pressure data of the working height of the area and vibration data of the safety belt during work, and inputs the aligned data into a safety belt intelligent dangerous state monitoring model for safety belt intelligent dangerous state monitoring. The safety belt intelligent dangerous state monitoring model comprises a perception field adaptive convolution module, a vibration frequency spectrum attention enhancement module, a cross-modal gating fusion layer, a time domain convolution gating recurrent unit, a multi-scale time sequence pooling layer, and a full-connection classifier. The present application can solve the problems of insufficient data alignment accuracy, limited feature extraction capability, low-efficiency multi-modal fusion, and incomplete time sequence dependence capture in the existing safety belt monitoring technology for aerial work.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent monitoring of safety belts, and particularly relates to a high-altitude operation safety belt dangerous state intelligent monitoring method and system and electronic equipment. BACKGROUND

[0002] In a high-altitude operation scene, a safety belt is core protective equipment for protecting the life safety of an operator, and real-time monitoring of the use state of the safety belt and dangerous early warning are directly related to the life safety of the operator. However, a traditional high-altitude operation safety supervision mode highly depends on manual on-site supervision, post-intervention and other extensive management means, has inherent defects such as limited monitoring range, lagging response and strong subjectivity, is difficult to cope with dynamic changes in personnel protection failure risks in a complex operation environment, and cannot meet the safety needs of modern high-altitude operation.

[0003] To solve the above problems, a monitoring system for the state of a safety belt has appeared in the prior art, for example, a safety monitoring system for a full-range safety belt proposed in Chinese Patent for Invention with publication number CN119832683A. The system integrates a controller, an environment monitoring module (temperature, humidity, wind speed, air pressure sensor), a position positioning module, a safety rope wearing monitoring module (pressure sensor, acceleration sensor), an audible and light alarm module and a remote data transmission module, realizes real-time collection and remote monitoring of the position of an operator, the wearing state of a safety rope and environmental parameters, and can trigger audible and light alarms when it is judged that the operator is in a protection failure state, thereby improving the safety supervision efficiency of high-altitude operation to a certain extent and reducing the protection failure phenomenon caused by carelessness and lucky psychology.

[0004] Although the prior art has made progress in hardware integration and basic monitoring functions, there are still significant limitations in the accuracy of data processing and dangerous state recognition, which are specifically as follows:

[0005] 1. Low-frequency environmental data (such as air pressure sampling rate 10Hz) and high-frequency dynamic data (such as vibration sampling rate 100Hz) need to be collected synchronously in high-altitude operation, and existing interpolation or time stamp alignment methods cannot realize accurate matching of the two, which easily destroys the time continuity of vibration data, resulting in deviation in dangerous state judgment.

[0006] 2. A traditional fixed-size convolution kernel is difficult to capture both the macro trend of air pressure data (such as height change) and the micro mutation of vibration data (such as instantaneous impact); and a conventional time domain convolution is not sensitive to frequency domain features, and cannot fully utilize frequency domain information (such as abnormal vibration of a specific frequency band) of vibration data to identify a dangerous state.

[0007] 3. The prior art mainly uses simple splicing to fuse multi-modal data such as air pressure and vibration, which cannot dynamically adjust the weight of different modal features according to real-time state, is difficult to play the synergistic advantage of multi-source data, and affects the robustness of dangerous state recognition.

[0008] 4. The traditional gated recurrent unit (GRU) only focuses on long-term time dependence and ignores local short-time features (such as the instantaneous mode of sudden impact), which cannot simultaneously consider short-time mutations and long-time trends, resulting in insufficient recognition accuracy for instantaneous dangers (such as sudden falls) and continuous abnormalities (such as equipment failures).

[0009] Therefore, how to break through the limitations of the prior art in data alignment, feature extraction, multi-modal fusion and time series modeling, and realize accurate and real-time monitoring of the dangerous state of the safety belt for aerial work, has become a technical problem to be solved. SUMMARY

[0010] The present application aims to solve the problems of insufficient data alignment accuracy, limited feature extraction capability, inefficient multi-modal fusion and incomplete time series dependence capture in the prior art safety belt monitoring technology for aerial work, and provides a safety belt dangerous state intelligent monitoring method, system and electronic equipment for aerial work, to realize accurate and real-time recognition of the dangerous state of the safety belt, and improve the safety of aerial work.

[0011] The present application is realized by the following technical solutions. A safety belt dangerous state intelligent monitoring method for aerial work, the steps are as follows:

[0012] S1, collecting air pressure data of the working height in the region and vibration data of the safety belt during work;

[0013] S2, labeling the collected air pressure data and vibration data based on the safety belt state of the aerial work safety scene knowledge base;

[0014] S3, aligning the air pressure data and the vibration data to form a two-dimensional aligned feature vector with a unified time reference;

[0015] S4, using the two-dimensional alignment feature vector as input, an intelligent safety belt dangerous state monitoring model is trained; the intelligent safety belt dangerous state monitoring model adopts an end-to-end time sequence classification architecture, the input layer receives the two-dimensional alignment feature vector, the first layer is a perception field adaptive convolution module, the convolution kernel width is dynamically adjusted according to the real-time air pressure gradient, and adaptive convolution operation is performed to extract convolution features; secondly, a vibration frequency spectrum attention enhancement module is accessed, the vibration frequency spectrum is calculated through short-time Fourier transform, the frequency domain attention weight is generated by using historical spectrum information, and the convolution features are enhanced accordingly; then, a cross-modal gate fusion layer is used to generate an air pressure gate signal based on the air pressure gradient value and the enhanced convolution features, and a vibration gate signal based on the vibration frequency spectrum and the enhanced convolution features, the air pressure related features and the vibration related features are dynamically weighted and fused to obtain fusion features; finally, a time domain convolution gate recurrent unit is used to extract local time sequence features from the fusion features, and a GRU unit is used to model long-range dependence, and an updated hidden state is obtained; a multi-scale time sequence pooling layer and a full connection classifier are used at the output end to output the safety belt state probability;

[0016] S5, the trained intelligent safety belt dangerous state monitoring model is integrated into an edge computing terminal for intelligent safety belt dangerous state monitoring.

[0017] Specifically, in step S2, the safety engineer combines the field monitoring video and the operation log to mark the safety belt state of each time window as four categories: safe steady state, instantaneous danger, continuous abnormality and environmental interference; the labeling tool uses a time sequence visualization platform, manually frames the vibration frequency spectrum abnormal interval and associates the air pressure fluctuation trend with the same timestamp, and finally generates a multi-dimensional label vector strictly synchronized with the sensor data.

[0018] Specifically, in step S3, the original air pressure data sequence is up-sampled to the same sampling rate as the vibration data sequence by using the cubic spline interpolation method.

[0019] Specifically, the process of dynamically adjusting the convolution kernel width according to the real-time air pressure gradient is as follows: using the reference kernel size as the basic width, multiplying the absolute value of the current air pressure gradient value by a preset kernel size adjustment coefficient, and performing a down rounding operation to obtain the dynamic convolution kernel width at the current time.

[0020] Specifically, the adaptive convolution operation is as follows: using the current time as the center, within the neighborhood range determined by the dynamic convolution kernel width, the two-dimensional alignment feature vector in the neighborhood is weighted and summed using learnable weight parameters, and a learnable bias term is added, then a linear rectifier unit is used for nonlinear transformation to obtain the convolution features at the current time.

[0021] Specifically, the processing process of the vibration frequency spectrum attention enhancement module is as follows:

[0022] At each time, a Hanning window is applied to the original vibration value within a fixed window length containing the time to reduce spectral leakage, then the short-time Fourier transform is performed on the data within the window, and the modulus value of the complex result is taken to obtain the spectral amplitude corresponding to each frequency point at the time;

[0023] For each frequency, the average value of the spectral amplitude in a historical time window is calculated, and then the attention weight value representing the importance of the frequency is calculated by using the frequency band weight vector and the bias term combined with the Sigmoid activation function;

[0024] First, find the dominant frequency with the maximum spectral amplitude at the current time, then obtain the attention weight value corresponding to the dominant frequency, and use the attention weight value to modulate the convolution features at the current time to strengthen the convolution features associated with the current dominant frequency, to obtain enhanced convolution features.

[0025] Specifically, the processing process of the cross-modal gating fusion layer is as follows:

[0026] The enhanced convolution features at the current time are spliced with the air pressure gradient value, then linearly transformed through the air pressure gating learnable weight matrix and the bias term, and then the Sigmoid activation function is applied to finally obtain the air pressure gating signal ranging from 0 to 1;

[0027] The enhanced convolution features at the current time are spliced with the maximum amplitude of the vibration spectrum, then linearly transformed through the vibration gating learnable weight matrix and the bias term, and then the Sigmoid activation function is applied to obtain the vibration gating signal;

[0028] The enhanced convolution features are linearly transformed by using the feature transformation weight matrix and passed through the linear rectifier unit to obtain the nonlinear enhanced vibration-related features, and the vibration gating signal is multiplied by the nonlinear enhanced vibration-related features; the air pressure gating signal value is multiplied by the enhanced convolution features to obtain the air pressure-related features, and finally the vibration-related features and the air pressure-related features processed by the gating signal are added to obtain the fusion features.

[0029] Specifically, the processing process of the time-domain convolution gating recurrent unit is as follows:

[0030] At each time, a fixed-size one-dimensional convolution kernel is used to perform convolution operation on the fusion features containing several time steps before and after the time centered at the time to obtain the local time sequence features representing the local time sequence pattern at the time;

[0031] The local time sequence features at the current time are spliced with the hidden state at the previous time, then linearly transformed through the update gate learnable weight matrix and the bias term, and then the Sigmoid activation function is applied to generate the update gate signal;

[0032] The local time sequence feature at the current moment is spliced with the hidden state at the previous moment, then linearly transformed through the reset gate learnable weight matrix and bias term, and a Sigmoid activation function is applied to generate a reset gate signal;

[0033] The hidden state at the previous moment is modulated by the reset gate signal, then the modulated historical state is spliced with the local time sequence feature at the current moment, and linearly transformed through the learnable weight matrix, and finally the hyperbolic tangent activation function is applied to obtain the candidate hidden state;

[0034] The hidden state at the previous moment and the current candidate hidden state are weighted and summed using the update gate signal as the weight, and the final output is the updated hidden state at the current moment.

[0035] The multi-scale time sequence pooling layer processing process is as follows:

[0036] At the end of the hidden state feature sequence, a maximum pooling operation is performed on each feature dimension, that is, the maximum value of each feature dimension in the time window is taken to obtain a short-time feature vector;

[0037] The hidden state features of all time steps are globally averaged pooled to obtain the average value of each feature dimension in the entire time period, then the cumulative sum sequence of the hidden state features with respect to the time step is calculated, and the maximum value of each feature dimension of the cumulative sum sequence is taken to obtain the peak value feature of the cumulative trend, finally, the global average feature vector and the cumulative peak value feature vector are added to obtain a long-time feature vector;

[0038] The extracted short-time feature vector and long-time feature vector are spliced to form a composite feature vector, then the composite feature vector is linearly transformed through the composite feature learnable weight matrix, and finally the normalized exponential function is applied to convert the transformation result into the probability distribution of each type of safety belt state.

[0039] Specifically, the risk perception loss function of the safety belt intelligent dangerous state monitoring model dynamically weights the loss term by using the historical confusion matrix, significantly increasing the high-risk error punishment for misjudging the dangerous state as a safe state, and the specific steps are as follows:

[0040] Define the risk coefficient weight:

[0041] ;

[0042] In the formula, The risk coefficient weight is used to weight the loss term of the th sample in the th class, and the greater the value indicates the stronger the punishment for misjudging the true class as the th class; is a risk reinforcement coefficient, controls the global strength of the misjudgment penalty; is a confusion matrix function, returns the frequency of being misjudged as the class in the training process, the frequency is dynamically updated based on historical batch data; is the true class label of the i-th sample; is the total number of classes; Calculate the risk-aware loss:

[0043]

[0044]

[0045] In the formula, represents the risk-aware loss, and the optimization target is to minimize the risk-aware loss; is the total number of samples in the training batch; is the sample index, and all samples in the batch are traversed, is the total number of classes; is the class index, and all classes are traversed, is an indicator function, and subscript i represents the association of sample x i and class c i When the true class of sample x i is c i , otherwise is the probability that the model predicts that sample x i belongs to class c i Subscript i represents the association of sample x i and class c i, satisfies is a logarithmic function, and the base is 10.

[0046] Specifically, in step S5, the air pressure data and the vibration data are monitored in real time; the double-flow asynchronous time alignment is performed according to step S3, the aligned real-time two-dimensional alignment feature vector is input into the safety belt intelligent dangerous state monitoring model, and the safety belt state probability is generated; when the instantaneous danger or continuous abnormal probability exceeds the threshold value, the sound and light alarm is triggered immediately, and the positioning information is sent to the management platform.

[0047] The application also provides a high-altitude operation safety belt dangerous state intelligent monitoring system, comprising:

[0048] A data acquisition module is used for acquiring air pressure data of the working height of the region and vibration data of the safety belt during work.​​​​​​​​​​​​​​​​​​

[0049] a data labeling module for labeling the collected air pressure data and vibration data based on the aerial work safety scene knowledge base according to the safety belt state;

[0050] a data preprocessing module for aligning the air pressure data and the vibration data to form a two-dimensional aligned feature vector with a unified time reference;

[0051] an edge computing terminal for deploying the trained safety belt intelligent dangerous state monitoring model to perform safety belt intelligent dangerous state monitoring.

[0052] The application further 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 aerial work safety belt dangerous state intelligent monitoring method.

[0053] The application has the following technical effects:

[0054] The cubic spline interpolation is adopted to realize accurate alignment of low-frequency air pressure data and high-frequency vibration data, avoid the time continuity of high-frequency vibration features from being destroyed, and lay a precise time sequence foundation for subsequent feature extraction; the perception field adaptive convolution module (dynamically adjusting the convolution kernel width) and the vibration spectrum attention enhancement module (strengthening the dangerous associated frequency band features) are combined to simultaneously capture the macro trend (such as height change) of the air pressure data and the micro mutation (such as instantaneous impact) of the vibration data, and solve the limitations of the traditional fixed convolution kernel in feature extraction; the cross-modal gating fusion dynamically weights the air pressure and vibration features, fully gives play to the synergistic advantages of multi-modal data, more accurately associates the coupling relationship between the two types of data compared with the simple splicing method, and significantly improves the recognition ability of instantaneous danger (such as sudden impact) and continuous abnormality (such as equipment failure).

[0055] The risk perception loss function dynamically weights the loss items based on the historical confusion matrix, greatly improves the punishment degree of the high-risk error that the dangerous state is misjudged as the safe state, effectively solves the model bias problem caused by the few dangerous samples and the class imbalance in the aerial work scene; the multi-scale time sequence pooling integrates the short-time maximum value pooling (capturing the instantaneous mutation within 10 seconds) and the long-time cumulative feature fusion (capturing the global trend), takes into account the short-time impact and the continuous abnormal mode, reduces the misjudgment caused by single-scale analysis, makes the model still output reliable results under complex environmental interference, and enhances the model stability and danger sensitivity.

[0056] The whole monitoring process (data acquisition, alignment, feature extraction, state judgment) has a delay of less than 50 milliseconds, which can respond to sudden dangers in high-altitude operation in real time (such as four scenes of safe steady state (normal operation), instantaneous danger (sudden impact), continuous abnormality (equipment failure), and environmental interference (wind noise and the like false alarm)); the combination of the edge computing terminal and the low-power Bluetooth transmission ensures that the sensor data are processed quickly and the alarm signal is triggered immediately (sound and light alarm + positioning information upload), so as to save the emergency response time for the operating personnel and directly improve the life safety guarantee efficiency.

[0057] Innovatively, through dynamic feature extraction, frequency domain information utilization and cross-modal dynamic fusion, the model can adapt to pressure fluctuations in different operation environments (such as high-low altitude difference) and vibration interference (such as wind noise and equipment shaking), and reduce false positives and false negatives caused by environmental factors; the deep fusion of multi-modal data (pressure + vibration) can more comprehensively reflect the operation state than single sensor monitoring, and improve the perception ability of the model to complex working conditions (such as severe weather and complex terrain). BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The method flowchart of the application.

[0059] Figure 2 The F1 score comparison column chart of different methods on four types of safety belt states.

[0060] Figure 3 The confusion matrix of the traditional double-flow splicing method.

[0061] Figure 4 The confusion matrix of the method of the application. DETAILED DESCRIPTION

[0062] The application will be further illustrated in detail below in combination with the drawings and examples.

[0063] Reference Figure 1 The application proposes a high-altitude operation safety belt danger state intelligent monitoring method, and the main steps are as follows:

[0064] S1, data acquisition: the data acquisition stage is realized through a multi-source sensor system installed at key nodes of the safety belt, specifically, a micro air pressure sensor (sampling rate 10 Hz) is used to continuously monitor the air pressure change of the operation height and record the low-frequency slowly changing air pressure data; a high-sensitivity gyroscope (sampling rate 100 Hz) is synchronously arranged to capture high-frequency vibration data such as instantaneous shaking and impact of the safety belt. The air pressure data reflect the macroscopic changes of the environment, such as height changes, and the vibration data record the microscopic motion state, such as equipment collision and personnel sliding. All sensors transmit the original data to the edge computing terminal in real time through low-power Bluetooth, and the data are stored in the form of time-stamped parallel streams, forming an initial monitoring sequence.

[0065] S2, data labeling: Based on the high-altitude operation safety scene knowledge base, the collected air pressure data and vibration data are labeled for safety belt state. By combining the field monitoring video and operation log, the safety belt state of each time window (10 seconds) is marked as four categories: safe steady state (normal operation), instantaneous danger (sudden impact), continuous anomaly (device failure), and environmental interference (wind noise and the like false alarm). The labeling tool adopts a time sequence visualization platform, manually frames the abnormal interval of the vibration spectrum and associates the air pressure fluctuation trend with the same timestamp, and finally generates a multi-dimensional label vector strictly synchronized with the sensor data.

[0066] S3, double-flow asynchronous data time sequence alignment: The sampling rate of the air pressure sensor data and the gyroscope data is significantly different. The air pressure data is low-frequency and slowly changing, while the vibration data is high-frequency and transient. Direct splicing will cause the time continuity of the high-frequency vibration characteristics to be destroyed, and the corresponding relationship between the two in time cannot be accurately reflected. The conventional technology usually directly interpolates the low-frequency data to the sampling rate of the high-frequency data, but this method may bring interpolation error, or simply aligns the timestamp, which cannot guarantee the time continuity of the vibration data. The embodiment adopts a cubic spline interpolation method to upsample the original air pressure data sequence to the same 100Hz sampling rate as the vibration data sequence. Specifically, cubic spline interpolation coefficients are calculated based on the original air pressure data points, and the cubic spline interpolation coefficients are used to generate the aligned air pressure value corresponding to each vibration sampling time point, so that the air pressure data and the vibration data are accurately aligned on the time axis and the time continuity of the vibration data is maintained. Finally, the aligned air pressure value at each time and the original vibration value at the corresponding time are spliced to form a two-dimensional aligned feature vector with a unified time reference. The specific steps are as follows:

[0067] Let the original air pressure data sequence be , and the vibration data sequence be , where is the length of the original air pressure data, and the length of the vibration data is , that is, the vibration sampling rate is 10 times that of the air pressure, are the original air pressure values at the 1st, 2nd, …, time, are the original vibration values at the 1st, 2nd, …, time, is the dimension space of the original air pressure data, is the dimension space of the vibration data. The aligned air pressure value is calculated by cubic spline interpolation:

[0068] ;

[0069] In the formula, is the aligned air pressure value at the time; represents a time index; represents a summation index, taking values from 0 to 3, indicating the order of the cubic spline polynomial; represents a cubic spline interpolation coefficient, obtained by solving the original air pressure data based on the cubic spline interpolation method; represents the latest original air pressure timestamp, i.e., the maximum original air pressure sampling time point less than or equal to ; represents the original air pressure value sampling interval, in seconds, for example, when the air pressure sampling rate is 10 Hz seconds;

[0070] The aligned air pressure value and the vibration value at the corresponding time are spliced to obtain an aligned feature vector, denoted as:

[0071] ;

[0072] In the formula, is a two-dimensional aligned feature vector; represents a vector splicing operation; is the original vibration value at the time; represents a two-dimensional real vector space, i.e., the feature dimension after splicing is 2.

[0073] S4, using the two-dimensional aligned feature vector as input, an intelligent safety belt dangerous state monitoring model is trained; the intelligent safety belt dangerous state monitoring model adopts an end-to-end time sequence classification architecture, the input layer receives the two-dimensional aligned feature vector output by step S3, the core structure includes four-level processing modules: the first layer is a perception field adaptive convolution module, then a vibration spectrum attention enhancement module is accessed, then a cross-modal gate fusion layer, and finally a time domain convolution gate recurrent unit is used to model the time sequence dependence; the output end passes through a multi-scale time sequence pooling layer, and a full-connection classifier outputs a safety belt state probability. All weight parameters are initialized with Xavier uniform distribution, and the bias term is initialized to zero.

[0074] S5, the trained intelligent safety belt dangerous state monitoring model is integrated into an edge computing terminal for intelligent safety belt dangerous state monitoring. Real-time air pressure data and vibration data are monitored; according to step S3, the double-flow asynchronous time sequence is aligned, the aligned real-time two-dimensional aligned feature vector is input into the intelligent safety belt dangerous state monitoring model, and four types of safety belt state probabilities are generated; when the instantaneous danger or continuous abnormal probability exceeds a threshold value, such as 0.92, an audible and visual alarm is triggered and positioning information is sent to the management platform. The whole process delay is less than 50 milliseconds, supporting real-time identification of four types of dangerous scenes: safe steady state (normal operation), instantaneous danger (sudden impact), continuous abnormality (equipment failure), and environmental interference (false alarm such as wind noise).

[0075] The air pressure data has macroscopic trend, and the vibration data has microscopic mutation characteristics, and the traditional fixed size convolution kernel is difficult to capture the characteristics of the two different scales at the same time, and the conventional processing method usually uses a single fixed size convolution kernel for feature extraction, but such a convolution kernel can only capture features of a specific scale, and cannot adapt to the characteristics of different signals.

[0076] The perception field adaptive convolution module of the embodiment dynamically adjusts the convolution kernel width according to the real-time air pressure gradient, performs adaptive convolution operation to extract convolution features of macroscopic trend or microscopic mutation, and the specific steps are as follows:

[0077] 1) Calculate the dynamic convolution kernel width

[0078] According to the gradient size of the aligned air pressure value at the current time, the width of the convolution kernel is dynamically adjusted, specifically, the reference kernel size is used as the basic width, the absolute value of the air pressure gradient value at the current time is multiplied by the preset kernel size adjustment coefficient, and the down rounding operation is performed, to obtain the dynamic convolution kernel width at the current time, so that the receptive field of the convolution kernel can adaptively change, which is expressed as:

[0079] ;

[0080] In the formula, is the dynamic convolution kernel width at the current time; is the reference kernel size, such as ; is the kernel size adjustment coefficient, such as , which controls the influence degree of gradient change on kernel size; represents the down rounding operation; is the air pressure gradient value at the current time, and the calculation method is expressed as ; is the aligned air pressure value at the current time.

[0081] 2) Perform adaptive convolution operation

[0082] Using the dynamic convolution kernel width, perform convolution operation on the spliced two-dimensional aligned feature vector, specifically, taking the current time as the center, within the neighborhood range determined by the dynamic convolution kernel width, using the learnable weight parameter to weight sum the two-dimensional aligned feature vectors in the neighborhood, and adding the learnable bias term, then through the linear rectifier unit to perform nonlinear transformation, to obtain the convolution feature at the current time, to extract local spatiotemporal features that can adapt to air pressure change mode, which is expressed as:

[0083] ;

[0084] In the formula, is​​​ Convolutional features extracted at each time step; It is a linear rectifier unit; It is a positive integer; Let be the learnable weight parameters corresponding to the i-th position; For the first Two-dimensional aligned feature vectors at each position; The convolution bias term is a learnable parameter.

[0085] It should be noted that when the pressure gradient value An increase indicates a dramatic change in air pressure, and the convolution kernel width... It automatically increases and expands the receptive field to capture long-range dependencies; when the air pressure changes are gradual, it uses a smaller convolutional kernel to focus on local features.

[0086] Dangerous conditions are usually manifested as abnormal vibrations in a specific frequency band. However, conventional time-domain convolution is not sensitive to the frequency domain features of vibration data and is difficult to effectively identify dangerous conditions. Conventional processing methods only perform convolution operations in the time domain, ignoring the importance of frequency domain features and failing to make full use of frequency domain information to identify dangerous conditions.

[0087] The vibration spectrum attention enhancement module in this embodiment calculates the vibration spectrum through short-time Fourier transform, generates frequency domain attention weights using historical spectrum information, and enhances convolutional features accordingly. The specific steps are as follows:

[0088] 1) Calculate the vibration spectrum

[0089] The vibration data undergoes time-frequency transformation to obtain spectral information. Specifically, at each time point, a Hanning window is applied to the original vibration values ​​within a fixed window length, including that time point, to reduce spectral leakage. Then, a short-time Fourier transform is performed on the data within the window, and the modulus of the complex result is taken to obtain the spectral amplitude corresponding to each frequency point at that time point, thus obtaining the frequency domain representation of the vibration data at each time point, expressed as:

[0090] ;

[0091] In the formula, for Time Frequency Spectral amplitude at a given location; frequency The value of is a positive integer; This is the index for sampling points within the Hanning window, with a value range of... arrive This indicates the location inside the Hanning window; For time index The original vibration value at the location; For the location of Hanning window The value at that location, This is used to reduce spectrum leakage; Here are the basis functions for the Fourier transform, where The imaginary unit; For example, the length of the window, Controlling spectral resolution; This indicates taking the modulus of a complex number to obtain the spectral amplitude.

[0092] 2) Calculate the frequency domain attention weights

[0093] For each frequency, the average spectral amplitude over a historical time window is calculated. Then, using a learnable band weight vector and a bias term, combined with a sigmoid activation function, an attention weight value representing the importance of that frequency is calculated, generating a weight vector between 0 and 1. A larger weight value indicates that the frequency band is more important in the recent past, expressed as:

[0094] ;

[0095] In the formula, For frequency The corresponding attention weight value ranges from (0, 1); Use the Sigmoid activation function; The frequency band weight vector consists of learnable parameters. superscript Indicates transpose; This indicates an average pooling operation that aggregates data over the time dimension. Indicates from arrive Frequency of time Spectral amplitude sequence, where For the length of the history window, e.g., ; For frequency The corresponding bias term is a learnable parameter.

[0096] 3) Enhance convolutional features

[0097] To enhance convolutional features in the time domain using spectral information, we first identify the dominant frequency with the largest spectral amplitude at the current moment. Then, we obtain the attention weight value corresponding to this dominant frequency. This attention weight value is then used to modulate the convolutional features at the current moment, thereby strengthening the convolutional feature representation associated with the current dominant frequency. This is expressed as follows:

[0098] ;

[0099] In the formula, for Time-enhanced convolutional features; for Convolutional features at time step; represents an element-wise multiplication operation; for the dominant frequency corresponding attention weight value; for moment spectrum The maximum amplitude frequency is calculated as .

[0100] It should be noted that when the spectral energy of a specific frequency band continues to rise, the specific frequency band is associated with a dangerous state, and the corresponding attention weight value increases, thereby enhancing the convolution feature expression related to the frequency band through feature modulation.

[0101] There is associated information between air pressure and vibration modalities, but simple feature splicing cannot effectively utilize this associated information, and the advantages of multi-modal data cannot be fully utilized. The conventional processing method usually simply splices the features of different modalities and directly inputs the model for processing, but this splicing method cannot dynamically adjust the importance of different modalities.

[0102] The cross-modality gating fusion layer of the embodiment generates an air pressure gating signal based on the air pressure gradient value and the enhanced convolution feature, generates a vibration gating signal based on the vibration spectrum and the enhanced convolution feature, dynamically weights and fuses the air pressure related features and the vibration related features, and obtains the fusion features. The specific steps are as follows:

[0103] 1) Calculate the air pressure gating signal

[0104] The gating signal that controls the importance of the air pressure feature is generated. Specifically, the enhanced convolution feature at the current moment is spliced with the air pressure gradient value, then linearly transformed through the air pressure gating learnable weight matrix and the bias term, and then the Sigmoid activation function is applied, and finally the air pressure gating signal ranging from 0 to 1 is obtained, which is represented as:

[0105] ;

[0106] In the formula, is the air pressure gating signal at the moment, the value range is (0, 1); is the air pressure gating learnable weight matrix, which controls the feature transformation; represents the vector splicing operation; is the air pressure gradient value at the moment, is the air pressure gating bias term, which is a learnable parameter. 2) Calculate the vibration gating signal

[0107]

[0108] ​​A gated signal controlling the importance of vibration features is generated. Specifically, the enhanced convolutional features at the current moment are concatenated with the maximum amplitude of the vibration spectrum. Then, a linear transformation is performed using the vibration-gated learnable weight matrix and bias term. Finally, the sigmoid activation function is applied to obtain the vibration-gated signal, expressed as:

[0109] ;

[0110] In the formula, for The vibration gating signal at time t, with a value range of (0, 1); For vibration-gated learnable weight matrix; for Vibration spectrum at any time The maximum amplitude, i.e. , characterizing the intensity of vibrational energy; This is a vibration-gated bias term, which is a learnable parameter.

[0111] 3) Execution Feature Fusion

[0112] The dual-modal features are fused using gated signals. Specifically, firstly, the enhanced convolutional features are linearly transformed using a feature transformation weight matrix and passed through a linear rectifier unit to obtain nonlinearly enhanced vibration-related features. The vibration-gated signal is then multiplied by the nonlinearly enhanced vibration-related features. The pressure-gated signal value is then multiplied by the enhanced convolutional features to obtain pressure-related features. Finally, the vibration-related features and pressure-related features processed from the gated signal are added together to obtain the fused features, represented as follows:

[0113] ;

[0114] In the formula, for Moment fusion features, superscript Indicates "integration"; The feature transformation weight matrix is ​​a learnable parameter that maps the input features to a new space.

[0115] It should be noted that when the air pressure changes significantly ( When (increases), Increase the importance of pressure-related characteristics; when vibration energy increases ( When (increases), Increase and strengthen the expression of vibration-related features.

[0116] It should also be noted that, This provides nonlinear feature enhancement and works in conjunction with vibration gating.

[0117] Traditional gated recurrent units (GRUs) neglect the extraction of local temporal features when processing time-series data, focusing only on long-range dependencies. The temporal convolutional gated recurrent unit of this invention first extracts local temporal features through one-dimensional convolution of the fused features, then inputs them into an improved GRU unit to model long-range dependencies, obtaining updated hidden states. The specific steps are as follows:

[0118] 1) Extracting local temporal features

[0119] Local temporal patterns are extracted from the fused feature sequence. Specifically, at each time step, a fixed-size one-dimensional convolutional kernel is used to perform convolution operations on the fused features, including those at several time steps before and after that time step, centered on that time step. This yields local temporal features representing the local temporal pattern at that time step, capturing the local dependencies of the fused features within a short time window, as shown below:

[0120] ;

[0121] In the formula, for Local temporal characteristics at any given moment; The convolution kernel weight matrix is ​​represented as One-dimensional convolution operation; arrive For time window The fusion feature sequence within; The kernel size is, for example, Control the local time range; is the convolution kernel weight matrix, which are learnable parameters.

[0122] 2) Calculate and update the gate signal

[0123] The local temporal features of the current time step are concatenated with the hidden state of the previous time step. Then, a linear transformation is performed through the learnable weight matrix and bias term of the update gate. Finally, the Sigmoid activation function is applied to generate the update gate signal. The resulting gate value controls the proportion of the transmission from the historical state to the new state, expressed as:

[0124] ;

[0125] In the formula, for The update gate signal at each time step, with a value range of (0, 1); To update the gate learnable weight matrix; This represents a vector concatenation operation; for The hidden state at any given moment; The gate bias term is a learnable parameter used to update the gate bias term.

[0126] 3) Calculate the reset gate signal

[0127] The local temporal features of the current time step are concatenated with the hidden state of the previous time step. Then, a linear transformation is performed using the learnable weight matrix and bias term through a reset gate. Finally, the Sigmoid activation function is applied to generate a reset gate signal. The resulting gate value controls the degree of influence of the hidden state of the previous time step on the calculation of the current candidate state, expressed as:

[0128] ;

[0129] In the formula, for The reset gate signal at time t, with a value range of (0, 1); To reset the gate learnable weight matrix; This is a learnable parameter used to reset the gate bias term.

[0130] 4) Calculate the candidate hidden state

[0131] The hidden state from the previous time step is modulated using a reset gate signal. Then, the modulated historical state is concatenated with the local temporal features of the current time step. A linear transformation is then performed using a learnable weight matrix, and finally, the hyperbolic tangent activation function is applied to obtain the candidate hidden state, represented as:

[0132] ;

[0133] In the formula, for The candidate hidden state at each moment; It is the hyperbolic tangent activation function; The weight matrix is ​​a learnable weight matrix; This represents element-wise multiplication. This indicates that historical information is filtered using the reset gate.

[0134] 5) Update hidden status

[0135] Using the update gate signal as weights, the hidden state from the previous time step and the current candidate hidden state are weighted and summed. The final output of the current hidden state retains some historical information while incorporating new information generated based on the current input and local features, balancing long-range dependencies and local patterns. This is represented as:

[0136] ;

[0137] In the formula, for The hidden state is constantly updated; To preserve historical information components; This is the newly added candidate information component.

[0138] It should be noted that the convolutional layer preferentially extracts local features within the time window , enhances the sensitivity to short-term patterns; reset gate signal controls the historical state of the current candidate state; update gate signal balances the contribution proportion of the previous hidden state and the candidate hidden state .

[0139] The aerial work safety belt state classification needs to capture both short-term mutations and long-term trends, but the traditional single pooling method is difficult to balance instantaneous danger signals and sustained abnormal states. The multi-scale temporal pooling layer of the present invention extracts short-term (near maximum) and long-term (global average + cumulative sum maximum) features from the hidden state sequence and splices and fuses them for final classification, with the specific steps as follows:

[0140] 1) Short-term feature extraction

[0141] In the last part of the hidden state feature sequence, maximum pooling operation is performed on each feature dimension, that is, the maximum value of each feature dimension in the time window is taken to obtain the short-term feature vector, which highlights the most significant abnormal or mutation pattern in the recent time window, denoted as:

[0142] ;

[0143] In the formula, denotes the short-term feature vector, which is used to represent the instantaneous mutation characteristics of the safety belt state; is the maximum pooling operation, which takes the maximum value in the time interval ; is the hidden state updated at the moment; is the total length of the time sequence; denotes the time window interval, covering the last 10 time steps, which is used to focus on short-term mutations.

[0144] 2) Long-term feature extraction

[0145] Capture the sustained trend characteristics of the safety belt state. Specifically, first, the global average pooling of the hidden state features of all time steps is performed to obtain the average value of each feature dimension in the entire time period, then the cumulative sum sequence of the hidden state features with time steps is calculated, and the maximum value of each feature dimension of the cumulative sum sequence is taken to obtain the peak value feature of the cumulative trend, finally, the global average feature vector and the cumulative peak value feature vector are added to obtain the long-term feature vector, which obtains a feature vector that fuses global statistical characteristics and cumulative abnormal trend information, representing the long-term evolution and sustained abnormality of the state, denoted as:

[0146] ;

[0147] In the formula, represents a long-time feature vector, used to represent the persistent trend feature of the safety belt state; is a global average pooling, which calculates the average value of all time step features; is a cumulative feature function, and the calculation method is represented as ; is a time sequence maximum value operation, which extracts the peak value of the cumulative feature in the whole time sequence.

[0148] 3) Feature fusion and classification

[0149] The state prediction is performed by integrating multi-scale information. Specifically, the extracted short-time feature vector and long-time feature vector are spliced to form a composite feature vector, then the composite feature vector is linearly transformed through a composite feature learnable weight matrix, and finally a normalized exponential function is applied to convert the transformation result into the probability distribution of each safety belt state, so that the model can utilize both the short-time features representing instantaneous dangerous signals and the long-time features representing persistent abnormal patterns to comprehensively judge the overall state of the safety belt and output the prediction probability, which is represented as:

[0150] ;

[0151] In the formula, represents the probability prediction vector of the safety belt state; is a normalized exponential function, which converts the output into a probability distribution; is a composite feature learnable weight matrix, which realizes the mapping from the feature space to the category space; represents a vector splicing operation, which connects and as a composite feature; is a short-time feature vector; is a long-time feature vector.

[0152] In the safety belt state classification of high-altitude work, the number of normal state samples is much larger than that of dangerous state samples, resulting in class imbalance. The traditional loss function is insufficient to punish the misjudgment of the dangerous state, which affects the recognition ability of the model for the dangerous state. The risk perception loss function of the embodiment dynamically weights the loss term by using the historical confusion matrix, significantly increases the high-risk error punishment intensity of misjudging the dangerous state as a safe state, and the specific steps are as follows:

[0153] 1) Define the risk coefficient weight

[0154] Based on the historical confusion matrix information collected during training, a risk coefficient weight is calculated for behaviors that misclassify the true category as another category. This results in high-risk errors, such as misclassifying a dangerous state as a safe state, being given a larger weight in the loss function and thus receiving a stronger penalty compared to other types of errors or low-frequency category misclassifications. This is expressed as:

[0155] ;

[0156] In the formula, Represents the risk coefficient weight, used for weighted calculation of the first... The sample at the th The larger the value of the loss term for each category, the better it represents the true category. Mistakenly identified as the first The stronger the punishment for each category; The risk enhancement coefficient controls the overall intensity of the penalty for misjudgment, such as... ; The confusion matrix function returns the true class during training. Misclassified as a category The frequency is dynamically updated based on historical batch data; For the first The true category label of each sample; This represents the total number of categories.

[0157] 2) Calculate perceived risk loss

[0158] Based on the standard cross-entropy loss, the loss term for each sample in each class is multiplied by the corresponding risk coefficient weight. Then, the weighted loss of all samples and all classes is averaged to construct a risk-aware loss function. This not only addresses the class imbalance problem but also strengthens the penalty for high-risk errors such as misclassifying dangerous states as safe states, thereby improving the model's sensitivity to identifying dangerous states. The risk-aware loss function is expressed as:

[0159] ;

[0160] In the formula, This represents the perceived risk loss, and the optimization objective is to minimize the perceived risk loss. This represents the total number of samples in the training batch. Using the sample index, iterate through all samples in the batch. ; Total number of categories; For category indexing, iterate through all categories. ; For indicator functions, subscript Indicates sample With category the association of the sample with the true class of the sample , otherwise ; is the probability that the model predicts the sample belongs to the class , subscript denotes the association of the sample with the class , satisfying and ; is a logarithmic function, with base 10.

[0161] The safety belt intelligent dangerous state monitoring model training adopts small batch random gradient descent, and 256 aligned time sequence samples are input per batch, each for 10 seconds. During the training process, the performance of the verification set is continuously monitored, and the weighted F1 score and risk perception loss value on the verification set are calculated every time a training cycle is completed. When the weighted F1 score of the verification set does not improve for 5 consecutive cycles, or the risk perception loss does not decrease for 5 consecutive cycles, the early stopping mechanism is triggered to save the current optimal model. At the same time, the maximum iteration cycle limit (200 cycles) is set to prevent overfitting. After the training is terminated, the model with the highest detection rate of dangerous states in the verification set is selected as the final safety belt intelligent dangerous state monitoring model.

[0162] In one embodiment, in order to evaluate the recognition ability of various monitoring methods for different states of the safety belt, the advantages of the method of the present application in detecting dangerous states are verified. The experiment compares the method of the present application with the traditional double-flow splicing method, the fixed convolution GRU (Gated Recurrent Unit) method and the vibration spectrum enhancement method, analyzes the recognition effect of four types of safety belt states (safe steady state, instantaneous danger, continuous abnormality and environmental interference), as shown in Figure 2 The method of the present application has a significant advantage in the two key dangerous states of instantaneous danger and continuous abnormality, and its column height is significantly higher than that of other methods, indicating the capture ability of the perception field adaptive convolution module for sudden impact and the reinforcement learning of the risk perception loss function for dangerous samples. It is worth noting that the traditional method performs the weakest in recognizing the instantaneous danger state, because the fixed convolution kernel is difficult to adapt to the characteristics of the sudden signal; and the method of the present application effectively captures the instantaneous abnormal jitter characteristics of the safety belt by dynamically adjusting the width of the convolution kernel.

[0163] In addition, reference is made to Figure 3 and Figure 4, by visualizing the classification error patterns, the decision characteristics of various methods are analyzed. The heat map uses color depth to represent the classification probability, and the diagonal element represents the correct recognition rate, and the non-diagonal element reveals the typical misjudgment mode. The heat map of the traditional method shows that there is a significant misjudgment flow (right upper dark block) from the dangerous state (especially the transient danger) to the safe steady state, and the high-risk error of "judging danger as safety" may cause serious consequences in practical application. The heat map of the method of the present application presents a clear diagonal advantage, especially in the color of the dangerous state column, which is the deepest, indicating that it effectively associates the coupling characteristics of pressure mutation and vibration spectrum through the cross-modal gating fusion layer. Especially in the transient danger row, the method of the present application greatly reduces the misjudgment to the safe steady state (lower left light block), indicating the specialized extraction capability of the multi-scale time sequence pooling layer to the short-time impact characteristics.

[0164] The application further provides a high-altitude operation safety belt dangerous state intelligent monitoring system, comprising:

[0165] A data acquisition module is configured to acquire pressure data of an operation height in a region and vibration data of a safety belt during operation;

[0166] A data labeling module is configured to label a safety belt state of the acquired pressure data and vibration data based on a high-altitude operation safety scene knowledge base;

[0167] A data preprocessing module is configured to align the pressure data and the vibration data to form a two-dimensional aligned feature vector with a unified time reference;

[0168] An edge computing terminal is configured to deploy a trained safety belt intelligent dangerous state monitoring model to perform safety belt intelligent dangerous state monitoring.

[0169] The application further 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 high-altitude operation safety belt dangerous state intelligent monitoring method.

[0170] The application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement each step of the high-altitude operation safety belt dangerous state intelligent monitoring method.

[0171] The above description only expresses the preferred embodiments of the application, and is not intended to limit the application in other forms. Any skilled person in the art can modify or reform the above disclosed content to equivalent embodiments with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the application, and according to the technical essence of the application, still belongs to the protection scope of the technical solution of the application.

Claims

1. A high-altitude operation safety belt danger state intelligent monitoring method, characterized by the steps of As follows: S1, collect the air pressure data of the working height of the region and the vibration data of the safety belt during work; S2, based on the high-altitude work safety scene knowledge base, the collected air pressure data and vibration data are labeled for safety belt state; S3, align the air pressure data and the vibration data to form a two-dimensional aligned feature vector with a unified time reference; S4, using the two-dimensional aligned feature vector as input, the safety belt intelligent dangerous state monitoring model is trained; the safety belt intelligent dangerous state monitoring model adopts an end-to-end time sequence classification architecture, the input layer receives the two-dimensional aligned feature vector, the first layer is a perception field adaptive convolution module, which dynamically adjusts the convolution kernel width according to the real-time air pressure gradient, performs adaptive convolution operation to extract convolution features; Secondly, the vibration frequency spectrum attention enhancement module is accessed, the vibration frequency spectrum is calculated by short-time Fourier transform, the frequency domain attention weight is generated using historical spectrum information, and the convolution features are enhanced accordingly; then, the cross-modal gating fusion layer is used to generate an air pressure gating signal based on the air pressure gradient value and the enhanced convolution features, and a vibration gating signal based on the vibration frequency spectrum and the enhanced convolution features, dynamically weighted fusion of air pressure related features and vibration related features to obtain fusion features; finally, the one-dimensional convolution is performed on the fusion features by the time domain convolution gating recurrent unit to extract local time sequence features, and the GRU unit is used to model the long-range dependence relationship to obtain the updated hidden state; the output end outputs the safety belt state probability through the multi-scale time sequence pooling layer and the full connection classifier; S5, the trained safety belt intelligent dangerous state monitoring model is integrated into the edge computing terminal for safety belt intelligent dangerous state monitoring.

2. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, In step S2, the safety engineer combines the field monitoring video and the work log to mark the safety belt state of each time window as four categories: safe steady state, instantaneous danger, continuous anomaly, and environmental interference; the annotation tool uses a time sequence visualization platform, manually frames the vibration frequency spectrum abnormal interval and associates the air pressure fluctuation trend with the same timestamp, and finally generates a multi-dimensional label vector strictly synchronized with the sensor data.

3. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, The process of dynamically adjusting the convolution kernel width according to the real-time air pressure gradient is as follows: using the reference kernel size as the basic width, multiplying the absolute value of the current time air pressure gradient value by the preset kernel size adjustment coefficient, and performing the down rounding operation to obtain the dynamic convolution kernel width at the current time; The adaptive convolution operation is as follows: taking the current time as the center, using the learnable weight parameter to weight and sum the two-dimensional aligned feature vectors within the neighborhood range determined by the dynamic convolution kernel width, and adding the learnable bias term, then performing nonlinear transformation through the linear rectifier unit to obtain the convolution features at the current time.

4. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, The processing process of the vibration frequency spectrum attention enhancement module is as follows: At each time, the original vibration value within a fixed window length containing the time is multiplied by a Hanning window to reduce spectral leakage, then the short-time Fourier transform is performed on the data within the window, and the modulus value of the complex result is taken to obtain the frequency spectrum amplitude value corresponding to each frequency point at the time; For each frequency, the average of its spectral amplitude in a historical time window is calculated, and then the attention weight value representing the importance of the frequency is calculated by combining the frequency band weight vector and the bias term with the Sigmoid activation function; First, find the dominant frequency with the largest spectral amplitude at the current time, then obtain the attention weight value corresponding to the dominant frequency, and use the attention weight value to modulate the convolutional features at the current time to strengthen the convolutional features associated with the current dominant frequency, thereby obtaining enhanced convolutional features.

5. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, The processing process of the cross-modal gating fusion layer is as follows: The enhanced convolutional features at the current time are concatenated with the air pressure gradient value, then linearly transformed through the air pressure gating learnable weight matrix and the bias term, and finally the Sigmoid activation function is applied to obtain the air pressure gating signal ranging from 0 to 1; The enhanced convolutional features at the current time are concatenated with the maximum amplitude of the vibration spectrum, then linearly transformed through the vibration gating learnable weight matrix and the bias term, and then the Sigmoid activation function is applied to obtain the vibration gating signal; The enhanced convolutional features are linearly transformed by the feature transformation weight matrix and passed through the linear rectifier unit to obtain the nonlinear enhanced vibration-related features, and the vibration gating signal is multiplied by the nonlinear enhanced vibration-related features; The air pressure gating signal value is multiplied by the enhanced convolutional features to obtain the air pressure-related features, and finally the vibration-related features and the air pressure-related features processed by the gating signal are added to obtain the fusion features.

6. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, The processing process of the time-domain convolution gating recurrent unit is as follows: At each time, a one-dimensional convolution kernel of a fixed size is used to perform convolution operation on the fusion features containing several time steps before and after the current time to obtain the local time sequence features representing the local time sequence pattern of the current time; The local time sequence features at the current time are concatenated with the hidden state of the previous time, then linearly transformed through the update gate learnable weight matrix and the bias term, and then the Sigmoid activation function is applied to generate the update gate signal; The local time sequence features at the current time are concatenated with the hidden state of the previous time, then linearly transformed through the reset gate learnable weight matrix and the bias term, and then the Sigmoid activation function is applied to generate the reset gate signal; The reset gate signal is used to modulate the hidden state of the previous time, then the modulated historical state is concatenated with the local time sequence features at the current time, and then linearly transformed through the learnable weight matrix, and finally the hyperbolic tangent activation function is applied to obtain the candidate hidden state; The update gate signal is used as the weight to weight the hidden state of the previous time and the current candidate hidden state, and finally the updated hidden state of the current time is output.

7. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, The processing process of the multi-scale time sequence pooling layer is as follows: In the last part of the hidden state feature sequence, a maximum pooling operation is performed on each feature dimension, that is, the maximum value of each feature dimension in the time window is taken to obtain a short-time feature vector; The hidden state features of all time steps are globally average-pooled to obtain the average value of each feature dimension in the entire time period, then the cumulative sum sequence of the hidden state features with time steps is calculated, and the maximum value of each feature dimension of the cumulative sum sequence is obtained to obtain the peak feature of the cumulative trend, finally, the global average feature vector and the cumulative peak feature vector are added to obtain the long-time feature vector; The extracted short-time feature vector and the long-time feature vector are spliced to form a composite feature vector, then the composite feature vector is linearly transformed through a composite feature learnable weight matrix, and finally, a normalization exponential function is applied to convert the transformation result into a probability distribution of each type of safety belt state.

8. The intelligent monitoring method for dangerous state of aerial work safety belt according to claim 1, characterized in that, The risk perception loss function of the safety belt intelligent dangerous state monitoring model dynamically weights the loss term by using the historical confusion matrix, significantly increasing the punishment intensity of the high-risk error of misjudging the dangerous state as the safe state, and the specific steps are as follows: Defining the dangerous coefficient weight: ； In the formula, Represents the risk coefficient weight, used for weighted calculation of the first... The sample at the th The larger the value of the loss term for each category, the better it represents the true category. Mistakenly identified as the first The stronger the punishment for each category; This is a risk enhancement coefficient used to control the overall intensity of penalties for misjudgments. The confusion matrix function returns the true class during training. Misclassified as a category The frequency is dynamically updated based on historical batch data; For the first The true category label of each sample; Calculating the risk perception loss: ； wherein, denotes the risk-aware loss, and the optimization objective is to minimize the risk-aware loss; is the total number of samples in the training batch; is the sample index, which iterates through all samples in the batch, ; is the total number of classes; is the class index, which iterates through all classes, ; is the indicator function, with subscript denoting the association of sample with class , when the true class of sample is , otherwise ; ; is the probability that model predicts sample belongs to class , with subscript denoting the association of sample with class , satisfying and ; is the logarithm function with base 10.

9. A system for implementing the intelligent monitoring method of dangerous states of aerial work safety belts according to any one of claims 1-8, characterized in that it comprises: Including: A data acquisition module for acquiring air pressure data of the working height in the region and vibration data of the safety belt during work; A data labeling module for labeling the safety belt state of the collected air pressure data and vibration data based on the high-altitude work safety scenario knowledge base; A data preprocessing module for aligning the air pressure data and the vibration data to form a two-dimensional aligned feature vector with a unified time reference; An edge computing terminal for deploying the trained safety belt intelligent dangerous state monitoring model to perform safety belt intelligent dangerous state monitoring.

10. An electronic device comprising a memory and a processor, the memory storing computer-executable instructions, characterized in that, The processor calls computer executable instructions to perform each step of the high-altitude work safety belt dangerous state intelligent monitoring method of any one of claims 1-8.

Images