A wind and overturning risk prediction method for presence devices combined with environmental parameters

Abstract

The present application relates to a kind of in presence equipment wind and overturning risk prediction method combined with environmental parameters, belong to building construction safety warning technical field.It includes the following steps: constructing the anti-wind overturning monitoring sample set comprising equipment type, risk category and data source label;By equipment type grouping numerical calibration, gust tilt mutation coupling, equivalent overturning moment reconstruction and foundation softening cumulative modeling, the original monitoring quantity is converted into physical enhanced time series input matrix with clear mechanical meaning;Constructing equipment adaptive time series risk classification model, fusing equipment type embedding, overturning driving force gated recurrent unit and risk evidence attention pooling module, and using category weighted loss to complete training;Finally, real-time acquisition of multi-source monitoring data, input trained model output risk level.The present application can effectively improve the response speed and identification accuracy of sudden overturning risk, realize the unified adaptive risk assessment of tower crane, construction elevator and other multiple equipment.

Description

Technical Field

[0001] This invention belongs to the field of construction safety early warning technology, specifically relating to a method for predicting the wind and overturning risks of on-site equipment by combining environmental parameters. Background Technology

[0002] In the construction industry, tower cranes and construction hoists are critical vertical transportation and lifting equipment, and their operational safety is directly related to the safety of personnel and property on construction sites. However, construction sites generally face the combined challenges of multiple environmental and working conditions, including strong winds, sudden gusts, heavy-load lifting, and changes in foundation bearing capacity. When the equipment is in a high, independent position or during heavy-load rotation, the overturning moment generated by wind loads and load eccentricity can easily exceed the overturning resistance of the foundation, leading to major accidents such as overall equipment instability or even overturning.

[0003] Existing technologies suffer from the following problems: Conventional monitoring methods treat wind speed, tilt angle, and load as independent input features, neglecting the mechanical coupling relationship between wind load, load eccentricity, and foundation weakening. When high-risk samples are insufficient, the model struggles to stably learn complex working conditions. Ordinary recurrent neural networks employ a uniform gating mechanism for all time steps. When faced with short-term strong gusts, historical stable states can easily dilute current sudden danger information, leading to a lag in the response to early signs of overturning risk. Different construction equipment exhibit significant differences in structural response scales. Existing methods either use a uniform model that ignores equipment characteristics or model each type of equipment separately, resulting in scattered training data and difficulty in achieving balance under conditions of scarce high-risk samples. Traditional classification methods rely solely on end-time features or simple time pooling for risk assessment, easily missing short-term danger signals appearing in the middle of the window. Furthermore, they lack temporal consistency constraints on continuous prediction results, and single-window noise can easily trigger false alarms. Summary of the Invention

[0004] To achieve the above objectives, the present invention employs the following technical solution: This invention provides a method for predicting the wind and overturning risks of on-site equipment by incorporating environmental parameters, comprising the following steps: S1. Obtain wind-resistant overturning monitoring samples and construct a training sample set containing equipment type labels, risk category labels, and data source labels; S2. Through numerical calibration by equipment type grouping, coupling of gust tilt change, reconstruction of equivalent overturning moment, cumulative modeling of foundation softening, and splicing of physical enhancement features, the original monitoring quantities are transformed into a physical enhancement time series input matrix with clear mechanical meaning. The physical enhancement time series input matrix is ​​then standardized channel by channel to obtain a standardized physical enhancement time series input matrix. S3. Construct an adaptive temporal risk classification model for equipment. The adaptive temporal risk classification model for equipment is trained by embedding equipment type, fusing physical enhancement features, gating recurrent units for overturning driving force, attention pooling of risk evidence, and category-weighted loss to achieve adaptive risk identification of equipment. S4. Continuously collect multi-source monitoring data of construction equipment, construct a standardized physical augmentation time-series input matrix, and input the standardized physical augmentation time-series input matrix into the trained equipment adaptive time-series risk classification model to obtain the risk level.

[0005] Furthermore, in step S1, the first Each time window serves as a sample for wind resistance and overturning monitoring. , This represents the sample index, with a value range of [value range missing]. , This represents the total number of wind resistance overturning monitoring samples, and each wind resistance overturning monitoring sample contains... Each time step Indicates the number of time steps within a single time window; the first Multi-source monitoring data for each wind resistance overturning monitoring sample includes continuous wind speed. Instantaneous gust wind speed Wind angle boom rotation angle Lifting weight Equipment tilt angle Foundation bearing capacity monitoring values and equipment operating status , This represents the time step index, with a value range of [value range missing]. Configure equipment type labels for each wind-resistant overturning monitoring sample. and risk category labels Record the data source label for each wind resistance overturning monitoring sample. .

[0006] Further, in step S2, data alignment, outlier handling, and grouping statistics are completed for the wind-induced overturning monitoring samples based on the equipment type labels, resulting in the basic monitoring time series matrix: For the first The timestamps of each sensor in the wind-resistant overturning monitoring sample are aligned to ensure that the continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle, foundation bearing capacity monitoring values, and equipment operating status correspond at the same time step; outliers exceeding the physical range of the sensors are clipped; and the original monitoring data after time alignment and outlier processing are spliced ​​together in a fixed channel order to obtain the basic monitoring time series matrix. Within the training sample set, the device type labels are statistically analyzed according to the device type labels. Feature channel mean and characteristic channel standard deviation .

[0007] Furthermore, in step S2, the enhancement degree of the instantaneous gust speed relative to the local gust background is coupled with the abrupt change intensity of the equipment tilt angle relative to the normal fluctuation scale of similar construction equipment to obtain the gust tilt abrupt change coupling characteristics: From the Instantaneous gust wind speed was read from a sample of wind resistance and overturning monitoring data. and equipment tilt angle Calculate the difference in device tilt angle between adjacent time steps. Based on the equipment type labels, the equipment tilt angle difference values ​​are grouped and statistically analyzed in the training set to obtain the equipment type labels. Mean difference in equipment tilt angle Standard deviation of equipment tilt angle difference ; Using the current time step as the endpoint, calculate the local average value of the instantaneous gust wind speed, and denot the local average gust wind speed as follows: ; The instantaneous gust speed at the current time step Local average gust wind speed The ratio was calculated to obtain the gust enhancement ratio. According to the first Equipment type label for individual wind resistance overturning monitoring samples Select the average difference of the corresponding equipment tilt angle. Standard deviation of equipment tilt angle difference Calculate the intensity of sudden changes in the tilt angle of the equipment. ; Increase the ratio of gusts and the intensity of sudden changes in equipment tilt angle By performing product coupling and compressing extreme values ​​using a hyperbolic tangent function, the coupling characteristics of gust tilt abrupt changes are obtained. .

[0008] Furthermore, in step S2, the multi-source monitoring data is uniformly transformed into driving force characteristics for overturning risk. Based on the overturning force rectangularity mechanism, the equivalent overturning moment is reconstructed, thereby obtaining the growth characteristics of the equivalent overturning moment: Read the first The number of wind resistance and overturning monitoring samples was in the first Continuous wind speed at each time step and instantaneous gust wind speed The maximum of the two values ​​is taken as the effective wind speed. ; Read the first The number of wind resistance and overturning monitoring samples was in the first Wind angle at each time step and boom rotation angle Calculate the relative angle difference between the two, and normalize the relative angle difference to... The effective windward projection degree is calculated within the range to obtain the wind direction and attitude projection factor. According to the effective wind speed Wind direction and attitude projection factor Comprehensive coefficient of pneumatic boom corresponding to equipment type Calculate the wind load moment term The comprehensive coefficient of the pneumatic boom refers to the equipment type label. The wind load calibration factor; Read the first The number of wind resistance and overturning monitoring samples was in the first Lifting weight at each time step and equipment tilt angle Calculate the eccentric moment term of the hoisting load. ; Read the first The number of wind resistance and overturning monitoring samples was in the first Foundation bearing capacity monitoring values ​​at each time step and the nominal reference value of foundation bearing capacity By comparison, the calculated foundation weakening moment term is denoted as... ; Wind load moment term 1. Lifting eccentricity moment item and foundation weakening moment term Adding them together, we get the equivalent overturning moment. Based on the equivalent overturning moment, the growth characteristics of the equivalent overturning moment were calculated. , , Indicates the first The number of wind resistance and overturning monitoring samples was in the first The equivalent overturning moment at each time step.

[0009] Furthermore, in step S2, a foundation softening cumulative index is constructed based on the extent of foundation bearing capacity loss within the time window, the proportion of time with insufficient foundation bearing capacity, and the longest continuous insufficient time. For the first For each time step of the wind-resistant overturning monitoring sample, the extent of foundation bearing capacity loss is calculated. The extent of the foundation bearing capacity loss is the nominal reference value of the foundation bearing capacity. Higher than the monitoring value of foundation bearing capacity Part; Statistics The number of time steps within the time window where the foundation bearing capacity of the wind-resistant overturning monitoring sample is insufficient. The number of time steps for insufficient foundation bearing capacity Divide by the time window length The percentage of time with insufficient foundation bearing capacity was obtained. The number of time steps where the foundation bearing capacity is insufficient is the total number of time steps in which the monitored foundation bearing capacity value is less than the nominal reference value of the foundation bearing capacity. Statistics The longest consecutive insufficient time step within the time window of each wind-resistant overturning monitoring sample. The longest continuous insufficient time step number is the longest time step during which the foundation bearing capacity monitoring value is continuously lower than the nominal reference value of the foundation bearing capacity. The average bearing capacity loss of the foundation within the calculation time window. The average bearing capacity deficit, the percentage of time with insufficient bearing capacity, and the number of longest consecutive insufficient time steps are combined to obtain the cumulative index of foundation softening. The formula is expressed as: ,in, Indicates the continuous deficiency penalty coefficient; and the cumulative index of foundation softening. Copy to the Each time step of the wind-resistant overturning monitoring sample forms a time series channel for the cumulative index of foundation softening.

[0010] Furthermore, in step S3, the device type embedding and physical enhancement feature fusion mapping are performed: Read the first Standardized physical augmentation time-series input matrix for one wind resistance overturning monitoring sample and device type label ; Create a device type embedded table Device type embedded table For trainable parameter matrices, based on device type labels Embedded table from device type Read device type embedding vector At each time step, the standardized physical augmentation feature vector is... Embedded vector with device type The device fusion feature vector is obtained by concatenating along the feature dimension; the device fusion feature vector is then input into the fusion mapping layer to obtain the fusion input vector. The fusion mapping layer consists of a fully connected layer, a layer normalization layer, and a hyperbolic tangent activation function in sequence.

[0011] Furthermore, in step S3, the overturning drive force gating cycle unit specifically includes: From the normalized physical enhancement time-series input matrix Read the standardized channel values ​​corresponding to the equivalent overturning moment and the standardized channel values ​​corresponding to the growth characteristics of the equivalent overturning moment; truncate the negative values ​​in the standardized equivalent overturning moment and the standardized equivalent overturning moment growth characteristics. The positive driving force input is obtained, and the positive driving force input is mapped to the overturning driving force activation value. ; fuse the input vector The hidden state vector of the previous time step and overturning driving force activation value The overturning driving force gated loop unit is used as a common input to obtain the hidden state vector at the current time step. The specific process is as follows: The overturning drive force gated loop unit is based on a fused input vector. and the hidden state vector of the previous time step The basic reset gate vector is calculated according to the mechanism of the standard gated loop unit. Preserve the gate vector with the basic state Activation value using overturning driving force Modulating these two basic gate vectors yields the reset gate vector. and state-preserving gate vector ; Based on the reset gate vector Filter the hidden state vector from the previous time step and combine it with the current fused input vector. Calculate the candidate hidden state vector : The hidden state vector of the previous time step With Reset Gate Vector Element-wise multiplication is used to filter historical states, and the filtered historical states are then fused with the current input vector. The inputs are combined into a fully connected layer and passed through a hyperbolic tangent activation function to obtain candidate hidden state vectors. ; Based on the state preservation gate vector The hidden state vector of the previous time step With candidate hidden state vector Perform weighted fusion to obtain the hidden state vector at the current time step. .

[0012] Furthermore, in step S3, the risk evidence attention pooling and risk category output specifically include: Read the first The hidden state vector sequence of each wind-resistant overturning monitoring sample at all time steps and overturning driving force activation value sequence Calculate the risk evidence score for each time step. The hidden state vector is input into a trainable linear scoring layer, and a weighted term of the overturning driving force activation value is superimposed to obtain the risk evidence score; the trainable linear scoring layer consists of a fully connected layer. The risk evidence weights are calculated by applying a normalized exponential function to the risk evidence scores for all time steps. The attention risk representation vector is obtained by weighting and summing the hidden state vector sequence according to the risk evidence weights. Max pooling is performed on the hidden state vector sequence along the time dimension to obtain the peak risk representation vector. , and the attention risk representation vector With peak risk representation vector By concatenating the vectors, a window-level risk representation vector is obtained. ; Window-level risk representation vector Inputting the fully connected classification layer yields the risk category probability distribution vector. The risk category corresponding to the highest predicted probability is output as the final risk category.

[0013] Furthermore, in step S3, the device adaptive time-series risk classification model is trained using category-weighted cross-entropy loss, and the final model parameters are selected using macro-average F1 score.

[0014] The advantages of this invention are: This invention constructs a physically guided feature reconstruction system. Through gust tilt abrupt change coupling, wind direction and attitude equivalent overturning moment reconstruction, and foundation softening cumulative modeling, it transforms raw multi-source monitoring data into enhanced features with clear mechanical meaning. This eliminates the need for the model to learn complex wind-load-foundation coupling relationships from massive high-risk samples. Furthermore, it constructs an overturning driving force-gated cyclic unit, directly injecting the equivalent overturning moment and its growth characteristics as physically gated signals into the cyclic network's state update process. When the overturning driving force increases, it automatically weakens the maintenance of historical stable states and strengthens the writing of current hazard information, thereby improving the response to sudden overturning risks. The system improves response speed; it proposes an equipment type embedding and adaptive fusion mechanism, mapping discrete equipment type labels to trainable dense vectors and fusing them with physical enhancement features at each time step, enabling a single model to handle tower cranes and construction hoists with different structural response characteristics, balancing sample utilization efficiency and risk identification specificity; it constructs an online inference framework that integrates risk evidence attention pooling and time consistency confirmation, using overturning driving force activation values ​​to guide attention weights to high-risk time steps, and suppressing single-window noise through a multi-window continuous confirmation mechanism, outputting interpretable risk source information while reducing false alarm rate. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0016] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 The data diagrams are for the construction process of the gust tilt change coupling feature of the present invention; wherein, (a) is a data diagram of instantaneous gust wind speed and its local average gust wind speed, (b) is a data diagram of equipment tilt angle difference and change intensity, and (c) is a data diagram of gust tilt change coupling feature. Figure 3 The output results of the online wind resistance and overturning risk assessment are shown in the figure; (a) is the sliding window risk probability output data figure, and (b) is the time consistency confirmation and alarm generation figure. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 In this embodiment, as Figure 1 As shown, this invention provides a method for predicting the wind and overturning risks of on-site equipment by combining environmental parameters. The specific steps include: S1. Sample Acquisition and Risk Category Labeling for Wind-Resistant Overturning Monitoring This invention uses tower cranes and construction hoists as monitoring objects for construction equipment, and acquires multi-source time-series data of construction equipment under different wind conditions, lifting conditions, attitude states, and foundation support states. The specific steps are as follows: 1) The wind-resistant overturning monitoring samples of the present invention can be obtained by on-site sensor acquisition, or by digital twin simulation based on finite element modeling. Alternatively, the samples obtained by on-site sensor acquisition and the samples obtained by digital twin simulation can be combined to form a training sample set.

[0019] In one implementation, measured wind-resistant overturning monitoring samples are acquired through on-site sensor data collection. Specifically, wind speed and direction sensors, tilt sensors, load weight sensors, boom angle sensors, and foundation bearing capacity monitoring devices are deployed in the construction equipment operating area to continuously collect data on continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle, foundation bearing capacity, and equipment operating status during the operation of the construction equipment.

[0020] In this implementation, the continuous wind speed is denoted as , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The average wind speed at each time step is used to characterize the intensity of a steady wind field; the instantaneous gust wind speed is denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The short-term peak wind speed at each time step is used to characterize the impact of sudden wind loads; the wind direction angle is denoted as... , This refers to the angle of the incoming airflow relative to a uniform reference direction; the boom rotation angle is denoted as... , This refers to the attitude angle of the main windward structure of a tower crane boom or construction hoist relative to the same reference direction; the lifting weight is denoted as... , This refers to the current lifting mass or equivalent lifting weight; the equipment tilt angle is denoted as... , This refers to the degree of tilt of construction equipment relative to the vertical direction; the foundation bearing capacity monitoring value is recorded as... , This refers to the monitoring results of the load-bearing capacity of the foundation or supporting area at the current time step; the equipment operating status is recorded as... , It refers to the operating status code of construction equipment at the current time step, such as stopped, running without load, rotating with load, lifting with load, or locked status.

[0021] In this implementation, the wind direction angle and boom rotation angle Use the same reference direction, such as true north or the coordinate system of the construction site. The positive direction of the axis is taken as Degrees. If the sensor output is in degrees, it should be converted to radians before being used in subsequent trigonometric function calculations to avoid calculation errors caused by inconsistent angle units.

[0022] In another implementation, simulated wind-resistant overturning monitoring samples are obtained through digital twin simulation based on finite element modeling. Specifically, a finite element digital twin model corresponding to the target construction equipment is established based on the structural parameters, material parameters, connection constraints, foundation support conditions, and hoisting arrangement of the construction equipment. A finite element digital twin model is a computational model used to simulate the structural response of construction equipment under the combined action of wind loads, hoisting loads, and foundation support conditions. It includes at least the tower or guide rail structure, the boom or main windward structure, the foundation support structure, the hoisting load application point, structural connection nodes, and boundary constraints.

[0023] In this implementation, the finite element digital twin model of the tower crane includes structural units such as the standard tower section, slewing bearing, boom, counterweight boom, and hook load; the finite element digital twin model of the construction hoist includes the guide rail frame, wall-mounted frame, cage, foundation support, and main connection nodes. Material parameters include elastic modulus, density, Poisson's ratio, and damping parameters; structural constraints include bottom foundation constraints, attachment constraints, slewing connection constraints, and load constraints; foundation support conditions can be characterized by the foundation spring stiffness, foundation rotational stiffness, or bearing capacity reduction factor. Through the above modeling method, the finite element digital twin model can output simulation responses such as equipment tilt angle, structural node displacement, foundation reaction force, and overall stability state.

[0024] In this implementation, simulation parameters are set in the finite element digital twin model to generate an input sequence consistent with the format of the field monitoring data. The simulation parameters include continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, foundation bearing capacity reduction state, and equipment operating status. Continuous wind speed and instantaneous gust wind speed can be generated according to the design wind speed level, historical wind speed statistics from the field, or a random wind field model; the wind direction angle can be generated from the design wind speed level, historical wind speed statistics from the field, or a random wind field model. Degree to Multiple directions can be set within the range of degrees; the boom rotation angle can be set according to the actual operating range of the equipment; the lifting weight can be set according to no-load, light load, medium load, rated load and over-limit approach state; the foundation bearing capacity reduction state can be set according to normal bearing, slightly insufficient, moderately insufficient and severely insufficient.

[0025] In another implementation, a layered combined sampling method can be used to generate simulation conditions. Specifically, multiple continuous wind speed levels and instantaneous gust wind speed levels are first set, and then different wind direction angles, boom rotation angles, load weights, and foundation bearing capacity reduction states are combined to form a set of conditions covering different levels of wind-induced overturning risk. For each simulation condition, the wind load is applied to the finite element digital twin model in a time sequence, the load weight is applied to the load application point or equivalent load node, and the foundation bearing capacity reduction state is converted into a reduction in foundation constraint stiffness or a reduction in the foundation bearing capacity monitoring value, thereby obtaining the equipment tilt angle response, structural displacement response, and foundation reaction force response under that condition.

[0026] In one embodiment, for example, the continuous wind speed can be set to... m / s m / s m / s and m / s, instantaneous gust wind speed is the same as the continuous wind speed Doubled Times, wind direction angle every Take one direction, and the working arm rotates at each angle. Take a certain orientation, and the lifting weight is taken as the rated load. , , , and The reduced bearing capacity of the foundation is taken as , , and By combining the above methods, it is possible to construct a large number of superimposed conditions of strong gusts, heavy loads, oblique winds, and weakened foundations that are difficult to collect safely on-site.

[0027] Furthermore, the simulation output of the finite element digital twin model is converted into a unified channel format for wind-resistant overturning monitoring samples. Specifically, the continuous wind speed in the simulation conditions is used as... The short-term peak wind speed in the simulation conditions is used as... The wind direction angle in the simulation condition is used as... The rotation angle of the working arm in the simulation condition is used as... The load weight in the simulation condition is used as The overall tilt response or top horizontal displacement of the equipment output from the finite element digital twin model is used as the conversion result. The bearing capacity of the foundation set in the simulation or the bearing capacity index obtained by back-deriving the foundation reaction force is used as... The simulation settings will be used as the operating status. Through this conversion, the simulated wind-resistant overturning monitoring samples and the measured wind-resistant overturning monitoring samples maintain consistency in data channels, number of time steps, and physical meaning.

[0028] In this implementation, when the finite element digital twin model directly outputs the device tilt angle, this device tilt angle is used as... When the finite element digital twin model outputs the top horizontal displacement but not the equipment tilt angle directly, the equivalent equipment tilt angle can be calculated based on the ratio of the top horizontal displacement to the effective equipment height. The effective equipment height refers to the vertical distance from the foundation support plane to the main response monitoring point. For example, for a tower crane, this can be the height from the bottom of the tower to the tower cap or slewing platform; for a construction hoist, it can be the height from the foundation to the top of the guide rail frame or the main stress position of the cage. This conversion ensures that the finite element simulation output and the on-site tilt sensor data have a consistent physical interpretation.

[0029] Furthermore, the finite element digital twin model is calibrated using on-site data to ensure comparability between the simulated wind-overturning monitoring samples and the measured wind-overturning monitoring samples. Specifically, several collected normal or low-risk operating conditions are selected as calibration conditions. The trends in equipment tilt angles, structural displacements, or foundation reaction forces obtained from on-site sensors are compared with the simulation output of the finite element digital twin model. The damping parameters, connection stiffness, foundation support stiffness, or wind load reduction factor in the finite element digital twin model are adjusted to ensure that the error between the simulation output and the measured response is within a preset range.

[0030] In practical implementation, the mean square error of the equipment tilt angle, the peak tilt angle error, and the response phase deviation can be used as calibration indicators. For example, when the maximum measured equipment tilt angle under low-risk on-site conditions is... The maximum equipment tilt angle output by the finite element digital twin model is [degree], while the maximum equipment tilt angle output by the finite element digital twin model is [degree]. When the simulation response amplitude is close to the field response, the structural connection stiffness can be appropriately increased or the wind load reduction factor can be adjusted. This calibration process can reduce the domain difference between simulation data and field data, and prevent the time-series risk classification model from overlearning unrealistic simulation response patterns.

[0031] 2) with the first Each time window is used as a wind resistance overturning monitoring sample, and the wind resistance overturning monitoring sample is denoted as . , It refers to a continuous multi-source time series data obtained by field sensor acquisition or digital twin simulation. This represents the sample index, with a value range of [value range missing]. , This represents the total number of wind resistance overturning monitoring samples. Each wind resistance overturning monitoring sample contains... Each time step This represents the time step index, with a value range of [value range missing]. , This represents the number of time steps within a single time window, for example, [can be taken as] When the sampling frequency is 1 Hz, take Corresponding to a 60-second monitoring window; when the sampling frequency is 2 Hz, take Corresponding to a 60-second monitoring window.

[0032] It should be noted that both the on-site sensor acquisition method and the digital twin simulation method require output wind-resistant overturning monitoring samples with the same time length and the same characteristic channel sequence. If the simulation time step is less than the on-site sampling time interval, the simulation output can be downsampled; if the simulation time step is greater than the on-site sampling time interval, it can be supplemented by interpolation so that the simulated wind-resistant overturning monitoring samples and the measured wind-resistant overturning monitoring samples do not need to distinguish the data source in subsequent processing.

[0033] 3) Configure an equipment type label for each wind-resistant overturning monitoring sample. The equipment type label is denoted as follows: , It refers to the first Each wind-resistant overturning monitoring sample corresponds to a category of building construction equipment. In one implementation, Corresponding tower crane, This corresponds to construction hoists. For measured wind-overturning monitoring samples, the equipment type label is determined by the on-site equipment register or equipment number; for simulated wind-overturning monitoring samples, the equipment type label is determined by the equipment type corresponding to the finite element digital twin model. The equipment type label is used for subsequent equipment type grouping statistics, equipment type embedding, and adaptive risk identification based on equipment type.

[0034] Furthermore, a risk category label is assigned to each wind-resistant overturning monitoring sample, and the risk category label is denoted as... , It refers to the first The true risk level of each wind-resistant overturning monitoring sample. In one implementation, the number of risk categories is denoted as... , The three risk categories are, in order, no risk, low risk, and high risk.

[0035] In practical implementation, for measured wind-resistant overturning monitoring samples, risk category labels can be determined based on on-site safety records, manual verification results, engineering threshold rules, or historical accident precursor records. The risk-free category corresponds to samples where the continuous wind speed, instantaneous gust wind speed, equipment tilt angle, and foundation bearing capacity monitoring values ​​are all within safe ranges; the low-risk category corresponds to samples with strong gusts, slight attitude changes, load eccentricity, or short-term insufficient foundation bearing capacity, but which have not yet reached the alarm or shutdown conditions; the high-risk category corresponds to samples where instantaneous gust impact, significant equipment tilt, large load eccentricity, or continuous insufficient foundation bearing capacity all occur simultaneously, reaching the standards for on-site alarm, operation restriction, or shutdown.

[0036] In practical implementation, for simulated wind-resistant overturning monitoring samples, the risk category label can be determined based on the stability index, maximum equipment tilt angle, foundation reaction force distribution, and overturning safety factor output by the finite element digital twin model. The overturning safety factor refers to the safety margin of the overturning resistance capacity of construction equipment relative to the overturning driving force, and can be calculated from the support reaction force, center of gravity offset, restoring moment, and overturning moment of the finite element digital twin model. When the overturning safety factor is greater than the first safety threshold and the maximum equipment tilt angle is less than the first tilt angle threshold, it is labeled as a no-risk category; when the overturning safety factor is between the first and second safety thresholds, or the maximum equipment tilt angle is between the first and second tilt angle thresholds, it is labeled as a low-risk category; when the overturning safety factor is less than the second safety threshold, or the maximum equipment tilt angle exceeds the second tilt angle threshold, or the finite element digital twin model shows local foundation heave, abnormal concentration of support reaction force, or overall instability trend, it is labeled as a high-risk category.

[0037] In one embodiment, for example, the first security threshold may be taken as follows: The second safety threshold can be taken as follows: The first tilt angle threshold can be taken as degrees, the second tilt angle threshold can be taken as Degree. If the overturning safety factor under a certain simulation condition is... The maximum equipment tilt angle is If the overturning safety factor is [value missing], then the simulated wind resistance overturning monitoring sample can be labeled as risk-free; if ... The maximum equipment tilt angle is If the degree is low, it can be marked as a low-risk category; if the overturning safety factor is If localized foundation uplift or abnormal concentration of support reaction forces occurs, it can be classified as a high-risk category. The above thresholds can be adjusted based on construction specifications, equipment models, on-site management requirements, or expert experience.

[0038] 4) Record the data source label for each wind-resistant overturning monitoring sample. The data source label is denoted as... , It refers to the first The source of the samples for wind-resistant overturning monitoring. In one implementation, This represents the measured wind resistance overturning monitoring samples obtained through on-site sensor acquisition. This represents simulated wind-resistant overturning monitoring samples obtained through digital twin simulation based on finite element method modeling. The data source label is not used as a risk category label, but rather for sample balancing, simulation sample weight control, or domain difference correction during the training phase.

[0039] In practical implementation, when the number of simulated wind-resistant overturning monitoring samples is much larger than the number of measured wind-resistant overturning monitoring samples, the ratio of measured to simulated wind-resistant overturning monitoring samples can be controlled in the training batch. For example, the proportion of measured wind-resistant overturning monitoring samples in each training batch can be adjusted. to Alternatively, a smaller loss weight can be set for the simulated wind-resistant overturning monitoring samples, while a larger loss weight can be set for the measured wind-resistant overturning monitoring samples. This allows the time-series risk classification model to better approximate the actual field distribution, while simultaneously using the simulated wind-resistant overturning monitoring samples to supplement extreme and dangerous working conditions.

[0040] 5) Combine the measured wind resistance overturning monitoring samples acquired by on-site sensors with the simulated wind resistance overturning monitoring samples acquired by digital twin simulation to form a training sample set. The training sample set is denoted as... , This refers to the dataset used to train a time-series risk classification model, denoted as... .in, This indicates the number of wind-resistant overturning monitoring samples in the training sample set. Indicates the first One wind-resistant overturning monitoring sample, Indicates the first The equipment type label corresponding to each wind-resistant overturning monitoring sample. Indicates the first Risk category label corresponding to each wind-resistant overturning monitoring sample Indicates the first Data source labels corresponding to each wind-resistant overturning monitoring sample.

[0041] It should be noted that the simulated wind-resistant overturning monitoring samples obtained through digital twin simulation are mainly used to supplement dangerous working conditions such as strong gusts, heavy loads, extreme wind directions, weakened foundations, and near-instability, which are difficult to collect safely on-site. The measured wind-resistant overturning monitoring samples obtained through on-site sensor acquisition are mainly used to maintain consistency between the training data and the actual construction site distribution. By unifying the two types of samples to the same data format and using them together for training, the coverage of high-risk samples can be expanded while ensuring the realism of the site, thereby improving the ability of the subsequent time-series risk classification model to identify low-frequency, high-risk overturning risks.

[0042] S2. Reconstruction and Enhancement of Physical Guiding Features for Wind-Resistant Overturning Assessment There are clear mechanical relationships among the original monitoring values ​​of construction equipment. Instantaneous gust wind speed will cause a sudden increase in wind load. The wind direction angle and the boom rotation angle together determine the effective windward projection. The load weight and the equipment tilt angle together determine the gravitational eccentricity effect. Insufficient foundation bearing capacity will weaken the foundation constraint capacity. If the continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle and foundation bearing capacity monitoring values ​​are used as independent input channels, the time series risk classification model needs to rely on a large number of samples to learn the complex coupling relationship on its own. When there are few high-risk samples, the identification is prone to instability.

[0043] This invention transforms raw monitoring data into a physically enhanced time-series input matrix with clear mechanical meaning through equipment type grouping numerical calibration, gust tilt abrupt change coupling, equivalent overturning moment reconstruction, foundation softening cumulative modeling, and physical enhancement feature stitching. The specific steps are as follows: S201, Equipment Type Grouping Numerical Calibration and Basic Monitoring Time Series Matrix Construction Tower cranes and construction hoists differ in their structural response dimensions. The same change in tilt angle may have different risk implications for tall tower cranes and rigid guide rail construction hoists; similarly, the same degree of foundation bearing capacity loss may correspond to different levels of danger under different foundation types.

[0044] To avoid the masking of true risk changes by differences in equipment type, this invention first completes data alignment, outlier handling, and grouped statistical preparation based on equipment type labels to obtain a basic monitoring time series matrix. The specific steps are as follows: 1) Regarding the first The timestamps of each sensor in the wind-resistant overturning monitoring sample are aligned so that the continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle, foundation bearing capacity monitoring value and equipment operating status have a corresponding relationship at the same time step.

[0045] In practical implementation, the data output from wind speed and direction sensors, tilt sensors, load weight sensors, and foundation bearing capacity monitoring devices can be uniformly resampled to the same sampling frequency, for example... Hertz or Hertz; for short-term missing values ​​between two valid sampling points, linear interpolation can be used to fill in the gaps; for time windows where consecutive missing values ​​exceed a preset missing threshold, such as consecutive missing values ​​exceeding [a certain threshold], [the missing values ​​can be filled in]. The time window can be removed or marked as invalid samples if the time is specified in seconds.

[0046] Furthermore, boundary clipping is performed on outliers that clearly exceed the physical range of the sensors. Specifically, continuous wind speed and instantaneous gust wind speed less than [a certain value] are considered outliers. Numerical clipping to Values ​​where the equipment tilt angle exceeds the sensor's range are clipped to the sensor's range boundary; the monitored value of the foundation bearing capacity is less than... Numerical clipping to This process can prevent abnormal amplification of subsequent physical enhancement features caused by single-point sensor failure.

[0047] 2) After time alignment and outlier processing, the original monitoring data are concatenated in a fixed channel order to obtain the basic monitoring time series matrix. Basic monitoring time series matrix It refers to the first The original multi-channel time series matrix formed after basic data cleaning of each wind resistance overturning monitoring sample, with dimensions of [dimensional value missing]. ;in, The number of original feature channels representing the basic monitoring time series matrix can be taken as follows in this invention. The eight original feature channels are, in order, continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle, foundation bearing capacity monitoring value, and equipment operating status.

[0048] 3) Within the training sample set, calculate the channel mean and channel standard deviation required for subsequent standardization according to the device type label. The device type label is... The characteristic channel mean is denoted as , This refers to the device type label. The mean vector of the training samples across each feature channel is used as the standardized center value for inputs of similar construction equipment; the equipment type label is... The characteristic channel standard deviation is denoted as , This refers to the device type label. The standard deviation vector of the training samples in each feature channel is used as a standardized scale value for the input of similar building construction equipment.

[0049] It should be noted that, and The statistics are calculated only from the training sample set and are directly reused in the validation, testing, and field inference phases, without being recalculated based on the samples to be evaluated. This design avoids changes in the relative strength of risk characteristics due to the numerical range of the samples to be evaluated, thereby improving the stability of risk assessment results.

[0050] S202, Construction of Gust Tilting Sudden Coupling Features Before construction equipment faces the risk of overturning due to wind, the common phenomenon is not a single increase in wind speed or a single change in tilt angle, but rather the simultaneous occurrence of a sudden increase in instantaneous gust wind speed and a sudden change in the equipment's tilt angle within a short period of time. If only instantaneous gust wind speed is used, ordinary strong winds can easily be misjudged as a risk of overturning; if only the equipment tilt angle difference is used, sensor noise or normal structural vibration can easily be misjudged as abnormal posture.

[0051] This invention couples the enhancement of instantaneous gust speed relative to the local gust background with the abrupt change intensity of the equipment tilt angle relative to the normal fluctuation scale of similar construction equipment to obtain the gust tilt abrupt change coupling characteristics. The specific steps are as follows: 1) From the first Instantaneous gust wind speed was read from a sample of wind resistance and overturning monitoring data. and equipment tilt angle Then, the equipment tilt angle difference between adjacent time steps is calculated, and the equipment tilt angle difference is denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The change in device tilt angle at each time step relative to the previous time step is used to characterize the rate of change in device attitude. In specific implementation, when season ;when At that time, it can be made This is to keep the time window length constant.

[0052] 2) Based on the equipment type label, perform training set grouping and statistical analysis on the equipment tilt angle difference values. Specifically, the equipment type label is... The mean difference in the equipment tilt angle is denoted as , This refers to the device type label. The average level of the equipment tilt angle difference values ​​in the training samples is used to characterize the typical attitude changes of similar construction equipment; the equipment type label is... The standard deviation of the equipment tilt angle difference is denoted as , This refers to the device type label. The fluctuation scale of the equipment tilt angle difference value in the training sample is used to measure the normal dispersion of the attitude change of similar construction equipment.

[0053] In one implementation, the statistic and The aim is to establish a benchmark for the attitude changes of similar construction equipment under normal operating conditions. This benchmark is used to measure whether new attitude changes constitute a "sudden change," that is, the degree of deviation from normal fluctuations. Therefore, it is advisable to prioritize the use of samples from the no-risk and low-risk categories for statistical analysis. and This reduces the contamination of normal fluctuation scales by drastic attitude changes in high-risk samples. If the sample size of a certain type of construction equipment is small, all training samples can be used for statistical analysis, and robust estimation can be performed using the median and interquartile range.

[0054] 3) Using the current time step as the endpoint, calculate the local average value of the instantaneous gust wind speed. Specifically, the length of the local average gust window is denoted as... The corresponding number of consecutive time steps involved in the local averaging calculation, for example, can be taken as... or .

[0055] In the specific implementation, in the first... At a time step, if Then use the first The time step to the 1 The local average gust speed is calculated as the instantaneous gust speed at each time step; if Then use the first The time step to the 1 The instantaneous gust wind speed at each time step is calculated as the local average gust wind speed. The local average gust wind speed is denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The background level of the gusts near the time step.

[0056] Furthermore, the instantaneous gust speed at the current time step... Local average gust wind speed The ratio was calculated to obtain the gust enhancement ratio. , This refers to the degree of enhancement of the current instantaneous gust speed relative to the background gust level over a recent period, expressed as... ;in, This represents the stability constant calculated using the gust ratio, used to prevent the denominator from being too large. For example, it is advisable .when Greater than When, it indicates that the instantaneous gust speed at the current time step is higher than the local gust background speed; when near This indicates that there has been no significant increase in gusts at the current time step.

[0057] 4) According to the first Equipment type label for individual wind resistance overturning monitoring samples Select the average difference of the corresponding equipment tilt angle. Standard deviation of equipment tilt angle difference Calculate the intensity of sudden changes in the tilt angle of the equipment. Among them, the intensity of sudden changes in equipment tilt angle. It refers to the first The number of wind resistance and overturning monitoring samples was in the first The degree to which the change in the equipment tilt angle at each time step deviates from the normal fluctuation range of similar construction equipment is calculated as follows: ; in, This represents the differential normalized stability constant of the equipment tilt angle, used to prevent the denominator from being... For example, it is advisable .

[0058] Sudden Intensity of Tilting Angle of Computing Device When taking the absolute value, it is used to capture both positive tilt abrupt changes and reverse swing abrupt changes simultaneously, avoiding sensitivity to changes in the device tilt angle in only one direction.

[0059] 5) Increase the gust ratio and the intensity of sudden changes in equipment tilt angle By performing product coupling and compressing extreme values ​​using a hyperbolic tangent function, the coupling characteristics of gust tilt abrupt changes are obtained. , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The coupling characteristics of gust tilt abrupt changes at each time step are used to characterize the degree to which the instantaneous increase in gust speed and the abrupt change in equipment tilt angle occur synchronously. The calculation method is expressed as follows: ; in, This represents the coupling strength adjustment coefficient, used to adjust the influence of the gust enhancement ratio and the abrupt change in equipment tilt angle on the coupling characteristics of the gust tilt change. For example, it can be taken as... to ; This represents the hyperbolic tangent function, used to limit the coupling characteristics of gust tilting abrupt changes within a stable numerical range.

[0060] In one embodiment, for example, the local average window length of the gust Pick If a certain tower crane is in the... The instantaneous gust speed at each time step is m / s, recently The local average gust wind speed at each time step is m / s, then the gust enhancement ratio is approximately If the equipment tilt angle difference value at this time step is standardized by tower crane grouping to obtain the equipment tilt angle change intensity... And the coupling strength adjustment coefficient The gust tilting abrupt change coupling characteristics are close to The output is close to This indicates a significant anomaly in the synchronization between gust winds and attitude changes at this time step. If the instantaneous gust wind speed increases but the intensity of the equipment tilt angle change remains close to [a certain value], this suggests an abnormality. If this is achieved, the coupling characteristics of gust tilt change will not increase significantly, thereby suppressing false alarms caused by ordinary wind speed fluctuations.

[0061] It should be noted that the gust tilt change coupling feature does not simply amplify the instantaneous gust wind speed, nor does it simply amplify the equipment tilt angle difference value. Instead, it emphasizes that the two types of hazard precursors appear synchronously near the same time step, thereby making it easier for the subsequent time series risk classification model to identify synchronous abnormal patterns before the occurrence of sudden overturning risk.

[0062] S203, Equivalent overturning moment reconstruction of wind direction and attitude coupling The risk of overturning of construction equipment is related to wind speed, wind direction, boom rotation angle, load weight, equipment tilt angle, and foundation bearing capacity. If conventional time series models only receive these raw monitoring values, they need to learn the mechanical relationship between wind load, load eccentricity, and foundation weakening from the samples themselves. This can easily lead to insufficient generalization when there are not enough high-risk samples.

[0063] Based on the mechanism of overturning force rectangularity, this invention reconstructs the equivalent overturning moment and unifies multi-source monitoring data into driving force characteristics oriented towards overturning risk. The specific steps are as follows: 1) Read the first The number of wind resistance and overturning monitoring samples was in the first Continuous wind speed at each time step and instantaneous gust wind speed The maximum of the two values ​​is taken as the effective wind speed, which is denoted as . , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The dominant wind speed input is used in the calculation of the overturning driving force at each time step.

[0064] In the specific implementation, let ;in, This represents the function that takes the maximum value. Through this processing, both the continuous wind field and the short-term peak gust can be included in the overturning risk assessment, and the peak gust will not be masked by the average wind speed.

[0065] It should be noted that since wind pressure is approximately related to the square of wind speed, the subsequent wind load moment term is calculated using the square of the effective wind speed, so that a more obvious overturning driving force response can be generated when the instantaneous gust wind speed increases.

[0066] 2) Read the first The number of wind resistance and overturning monitoring samples was in the first Wind angle at each time step and boom rotation angle Calculate the relative angle difference between the two.

[0067] In practical implementation, to avoid angle crossing degree or When a jump occurs, the relative angle difference is first normalized to... Within the range, the effective windward projection degree is then calculated to obtain the wind direction and attitude projection factor, which is denoted as . , This refers to the windward projection intensity formed by the combined direction of the incoming wind and the current posture of the construction equipment, and is calculated as follows: ; in, The device type label is The wind direction projection nonlinearity index is used to adjust the intensity of the influence of oblique winds on the windward projection of different construction equipment. For example, the wind direction projection nonlinearity index of a tower crane can be taken as... The nonlinear index of the wind direction projection of the construction hoist can be taken as... When the direction of the incoming wind is close to the direction of the main windward structure, the wind direction attitude projection factor is larger; when the direction of the incoming wind is nearly perpendicular to the direction of the main windward structure, the wind direction attitude projection factor is smaller.

[0068] 3) Based on effective wind speed Wind direction and attitude projection factor The aerodynamic boom comprehensive coefficient corresponding to the equipment type is used to calculate the wind load moment term. The aerodynamic boom comprehensive coefficient is denoted as... , This refers to the device type label. The wind load calibration factor is used to comprehensively characterize the influence of windward area, aerodynamic drag coefficient, and equivalent lever arm on the overturning driving force; the wind load moment term is denoted as... , This refers to the overturning driving component caused by wind load, and its calculation method is expressed as follows: .

[0069] In practical implementation, if all input features have already undergone dimensionless normalization, It can be used as a dimensionless engineering calibration coefficient; if the input features retain the true engineering dimensions, It is necessary to uniformly calibrate based on air density, windward area, drag coefficient, and lever arm to ensure that the wind load moment term and the hoisting eccentric moment term have consistent moment scales.

[0070] 4) Read the first The number of wind resistance and overturning monitoring samples was in the first Lifting weight at each time step and equipment tilt angle Calculate the lifting eccentric moment term, the calculation method is expressed as follows: .in, Indicates the lifting arm coefficient. This refers to the device type label. The load eccentricity calibration coefficient is used to characterize the ability of the load weight to generate an overturning moment when the equipment is tilted. This indicates the eccentric moment term during hoisting. It refers to the gravity eccentric drive component generated by the combined effect of the load weight and the equipment tilt angle.

[0071] In practical implementation, the device tilt angle The sine calculation needs to be performed in radians. When the equipment tilt angle is small, the eccentric moment term increases approximately linearly with the equipment tilt angle; when the load weight is large and the equipment tilt angle increases, the eccentric moment term increases synchronously, which can reflect the mechanical relationship that "heavy load tilting is more dangerous".

[0072] 5) Read the first The number of wind resistance and overturning monitoring samples was in the first Foundation bearing capacity monitoring values ​​at each time step and the nominal reference value of foundation bearing capacity Comparison. Among them, the nominal reference value of foundation bearing capacity. It refers to the reference benchmark used to determine whether the foundation support capacity meets the safety requirements, and can be determined by construction specifications, design documents, site survey reports or on-site acceptance records.

[0073] In one implementation, the ground bearing capacity monitoring value Lower than the nominal reference value of foundation bearing capacity At that time, the extent of foundation bearing capacity loss is calculated, representing the absolute degree of insufficient foundation support capacity at the current time step; when the foundation bearing capacity monitoring value Not lower than the nominal reference value of foundation bearing capacity At that time, the extent of the foundation bearing capacity loss was taken as... .

[0074] In this implementation, the foundation weakening moment term is further calculated based on the extent of foundation bearing capacity loss. The foundation weakening moment term is denoted as... , This refers to the additional overturning driving component introduced due to the decrease in foundation constraint capacity caused by insufficient foundation bearing capacity. The calculation method is expressed as follows: ; in, The device type label is The foundation weakening contribution coefficient is used to characterize the impact of insufficient foundation bearing capacity on the overturning risk of different construction equipment. Its preset value depends on the type of equipment. Sensitivity to foundation bearing capacity defects, for example: for construction hoists ( Due to its good overall structural integrity, the wall-mounted frame is quite sensitive to foundation settlement, and its contribution factor to foundation weakening is relatively high. Larger values ​​can be taken, for example For tower cranes ( Its independent height is high, the requirements for the overturning moment of the foundation are clearly defined, and its foundation weakening contribution coefficient is... A moderate value can be chosen, for example .

[0075] It should be noted that in calculating the foundation weakening moment term... In this case, by taking the positive part, the additional overturning effect is only introduced when the monitored value of the foundation bearing capacity is lower than the safety benchmark, so as to avoid the normal foundation bearing capacity state being erroneously amplified.

[0076] 6) Wind load moment term 1. Lifting eccentricity moment item and foundation weakening moment term Adding them together, we get the equivalent overturning moment. ,Right now Equivalent overturning moment It refers to the first The number of wind resistance and overturning monitoring samples was in the first The overturning driving force characteristics at each time step are formed by the combined effects of wind load, eccentric load, and insufficient foundation bearing capacity.

[0077] Furthermore, the growth characteristics of the equivalent overturning moment were calculated. , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The positive increase in the equivalent overturning moment at each time step relative to the previous time step is used to characterize whether the overturning driving force is increasing rapidly. In specific implementation, when season ;when season ;in, It refers to the first The number of wind resistance and overturning monitoring samples was in the first The overturning driving force characteristics at each time step are formed by the combined effects of wind load, eccentric load, and insufficient foundation bearing capacity.

[0078] It should be noted that by retaining only the positive growth rate, the subsequent time series risk classification model can focus on the time step when the risk driver rises rapidly, rather than mistaking the slow decline of risk as a process of increasing danger.

[0079] In one embodiment, for example, if all physical inputs have undergone dimensionless normalization, the tower crane can be... , , , Construction hoists can be used , , , The above values ​​can be re-determined through finite element digital twin simulation results, historical field data, or calibration by engineering experts, and are not intended to be fixed.

[0080] It should also be noted that the equivalent overturning moment transforms the monitored values ​​of continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle, and foundation bearing capacity into a single overturning driving force characteristic with clear mechanical meaning; the growth characteristic of the equivalent overturning moment further highlights the rapid increase process of the dangerous driving force. The combination of the two can reduce the dependence of subsequent time-series risk classification models on large-scale high-risk samples.

[0081] It should also be noted that in the model input construction of the present invention, all original monitoring quantities will be standardized in the subsequent step (S205) to eliminate dimensional differences. Therefore, the torque terms constructed in this step exist as risk characteristics, rather than torque values ​​in strict physical units.

[0082] S204, Construction of Accumulated Risk Characteristics of Time-Series Decrease in Foundation Bearing Capacity The impact of insufficient foundation bearing capacity on the risk of overturning of construction equipment depends not only on the extent of the defect at a certain time step, but also on the duration and continuity of the defect. The risks associated with short-term monitoring fluctuations are different from those associated with long-term foundation softening. If only the foundation bearing capacity monitoring value at a single time step is used, it is difficult to characterize the cumulative impact of continuous deterioration of foundation performance.

[0083] This invention constructs a foundation softening cumulative index based on the extent of foundation bearing capacity loss within a time window, the proportion of time with insufficient foundation bearing capacity, and the longest continuous insufficient time. The specific steps are as follows: 1) Regarding the first For each time step of the wind-resistant overturning monitoring sample, the extent of foundation bearing capacity loss is calculated. Degree of foundation bearing capacity loss This refers to the nominal reference value of the foundation bearing capacity. Higher than the monitoring value of foundation bearing capacity This part is used to characterize the degree of inadequacy of the foundation support capacity at the current time step, and is calculated as follows: .when hour, When it is a positive value; hour, for This treatment ensures that time steps where the foundation bearing capacity exceeds the safety benchmark will not offset the extent of damage at dangerous time steps.

[0084] 2) Statistics of the first The number of time steps within the time window where the wind-resistant overturning monitoring sample is located where the foundation bearing capacity is insufficient is recorded as follows: , It means satisfying The total number of time steps is used to characterize the duration of insufficient foundation bearing capacity.

[0085] Furthermore, the number of time steps for insufficient foundation bearing capacity will be determined. Divide by the time window length The percentage of time with insufficient foundation bearing capacity was obtained. , This refers to the proportion of time steps within a time window where the foundation bearing capacity is insufficient, and the calculation method is expressed as follows: The longer the period of insufficient foundation bearing capacity, the longer the foundation's insufficient support capacity persists.

[0086] 3) Statistics of the first The longest consecutive insufficient time step number within the time window of each wind-resistant overturning monitoring sample is denoted as . , It refers to the longest time step during which the monitored value of the foundation bearing capacity is continuously lower than the nominal reference value of the foundation bearing capacity, and is used to characterize whether the weakening state of the foundation exists continuously.

[0087] In practical implementation, it can be derived from... arrive Gradually scan the foundation bearing capacity monitoring values, when Increment the current consecutive count. ,when The current consecutive count is reset to zero, and the maximum consecutive count during the scanning process is recorded as the maximum consecutive count. .

[0088] 4) Calculate the average bearing capacity deficit within the time window. The average bearing capacity deficit is denoted as... , It refers to the first The average level of foundation bearing capacity loss of each wind-resistant overturning monitoring sample over the entire time window.

[0089] Furthermore, by combining the average bearing capacity deficit, the proportion of time with insufficient bearing capacity, and the number of longest consecutive insufficient time steps, a cumulative ground softening index is obtained. , It refers to the first The cumulative ground softening index of each wind-resistant overturning monitoring sample is used to characterize the overall degree of continuous deterioration of the foundation bearing capacity within a time window, and is calculated as follows: ; in, This represents the continuous deficiency penalty coefficient, used to adjust the intensity of the influence of the longest continuous deficiency time on the cumulative index of foundation softening. For example, it can be taken as... to When two time windows have the same average foundation bearing capacity deficit and the same percentage of time with insufficient foundation bearing capacity, the time window with the longer continuous period of insufficient bearing capacity has a higher foundation softening accumulation index, which is more in line with the risk accumulation law brought about by continuous foundation weakening.

[0090] In one embodiment, for example, a time window includes Each time step, of which If the monitored value of the foundation bearing capacity at any given time step is lower than the nominal reference value of the foundation bearing capacity, then the percentage of time with insufficient foundation bearing capacity is: If the average foundation bearing capacity loss is The longest consecutive insufficient time step number is Consecutive Insufficient Penalty Coefficient The cumulative index of foundation softening is: If another time window also has There are 100 insufficient time steps, but the longest consecutive number of insufficient time steps is only 100. If the cumulative index of foundation softening is low, it indicates that the window is more likely to be intermittent fluctuations rather than continuous foundation weakening.

[0091] 5) The cumulative index of foundation softening Copy to the For each time step of the wind-resistant overturning monitoring sample, a time series channel for the cumulative foundation softening index is formed. The cumulative foundation softening index time series channel is a constant value within the same time window, used to provide window-level foundation deterioration information to subsequent time series risk classification models. This method does not disrupt the original time step structure, while allowing each time step's risk assessment to be aware of the foundation softening degree within the current time window.

[0092] In practical implementation, the cumulative index of foundation softening It is a description of the whole A single numerical value representing the overall foundation weakening situation over a long time window is used to ensure that subsequent time-series models are aware of this window-level information at each time step. This scalar value is then used to represent this overall foundation weakening situation. As a new feature channel, and all time steps within that channel ( arrive The value of ) is assigned to For example, if a sample containing 60 time steps is calculated to... The newly constructed cumulative index channel for foundation softening is a sequence of length 60, where each element has a value of 0. This way, when the model makes a judgment at each time step, it can refer to the overall weakening overview of the entire window.

[0093] It should be noted that the foundation softening cumulative index considers the intensity, duration, and continuity of foundation bearing capacity loss simultaneously, which differs from the processing method that only uses the foundation bearing capacity monitoring value of a single time step. This can improve the ability to identify the slow cumulative overturning risk caused by continuous foundation bearing capacity deficiency.

[0094] S205, Physical Enhancement Timing Input Matrix Generation and Device Type Grouping Standardization After constructing the coupling characteristics of gust tilting abrupt changes, equivalent overturning moment, equivalent overturning moment growth characteristics, and foundation softening cumulative index, this invention splices the original monitoring channel with the physical enhancement channel to obtain the physical enhancement time-series input matrix for the time-series risk classification model. The specific steps are as follows: 1) Couple the characteristics of sudden gust tilting Equivalent overturning moment Equivalent overturning moment growth characteristics Time series channel of foundation softening cumulative index and foundation monitoring time series matrix By concatenating along the feature channel direction, the physical augmentation temporal input matrix is ​​obtained. Physically enhanced temporal input matrix It refers to the first The model input matrix formed by a number of wind-resistant overturning monitoring samples after physical guided feature reconstruction and enhancement, with dimensions of [missing information]. ;in, This represents the total number of feature channels in the physical enhancement timing input matrix.

[0095] In one embodiment, as an example, if the basic monitoring time series matrix contains The original feature channels were expanded to include gust tilt abrupt change coupling features, equivalent overturning moment, equivalent overturning moment growth features, and foundation softening cumulative index. For each physical enhancement channel, the physical enhancement timing input matrix is... Total number of feature channels for .

[0096] 2) According to the first Equipment type label for individual wind resistance overturning monitoring samples Select the corresponding feature channel mean and characteristic channel standard deviation For the physical augmentation timing input matrix Normalization is performed channel by channel to obtain the normalized physical augmentation timing input matrix. Standardized physical augmentation timing input matrix This refers to the physical augmentation timing input matrix after being standardized by device type grouping, with dimensions of [missing information]. .

[0097] In specific implementation, for the first For each feature channel, the mean and standard deviation of similar construction equipment channels saved during the training phase are used to perform subtraction of the mean and division by the standard deviation; if the standard deviation of a certain channel is... or close to Then add the standardized stability constant to the denominator. For example, it is advisable Discrete channels such as equipment operating status can be standardized using integer encoding, or they can be encoded using one-hot encoding and then spliced ​​with other channels.

[0098] In one embodiment, as an example, regarding the device operating status For processing such discrete channels, if integer encoding followed by standardization is used—for example, encoding states like shutdown, no-load, and loaded as 0, 1, and 2 respectively, and then using these integers as the channel value, calculating the mean and standard deviation together with other continuous value channels, and then standardizing—or if one-hot encoding is used followed by concatenation with other channels, for example, if there are 5 operating states, then... Converted into a 5-dimensional one-hot vector, the basic monitoring time series matrix is ​​now... The number of channels is no longer 8, but has become These unique heat vector channels, which consist of 0 / 1, no longer participate in the normalization process and are directly spliced ​​with other continuous channels after normalization.

[0099] 3) Retain the equivalent overturning moment before standardization. and the growth characteristics of equivalent overturning moment The channel position index is used by the subsequent overturning drive force gating loop unit to read the corresponding physical enhancement channel and convert the equivalent overturning moment and its growth characteristics into overturning drive force activation values. By retaining the channel position index, errors in reading physical gating signals due to changes in channel order during model implementation can be avoided.

[0100] It should be noted that the physical augmentation time series input matrix retains the original monitoring sequence structure in the time dimension, and embeds information on gust changes, equipment attitude response, wind direction attitude coupling, load eccentricity and foundation softening in the channel dimension. Compared with only inputting the original monitoring data, the physical augmentation time series input matrix can provide a clearer expression of the overturning risk precursors to the time series risk classification model.

[0101] S3, Equipment Adaptive Temporal Risk Classification Model Training The standardized physical augmentation time-series input matrix already contains physical augmentation features related to wind overturning risk, but the risk response patterns of tower cranes and construction hoists are still not entirely the same. Using completely uniform time-series classification parameters can easily weaken the differences between equipment types; training a separate model for each type of construction equipment would reduce sample utilization efficiency.

[0102] This invention constructs an adaptive temporal risk classification model for equipment based on overturning driving force perception. Through equipment type embedding, physical enhancement feature fusion, overturning driving force-gated recurrent units, risk evidence attention pooling, and category-weighted loss training, it achieves adaptive equipment risk identification under a unified model. The specific steps are as follows: S301, Device Type Embedding and Physical Enhancement Feature Fusion Mapping Discrete equipment type labels cannot directly participate in the continuous feature calculation of a neural network. This invention maps equipment type labels to trainable dense vectors and replicates these vectors to each time step, fusing them with a standardized physical augmentation temporal input matrix. This allows the temporal risk classification model to simultaneously perceive the construction equipment type and the current physical augmentation features at each time step. The specific steps are as follows: 1) Read the first Standardized physical augmentation time-series input matrix for one wind resistance overturning monitoring sample and device type label Among them, the normalized physical enhancement time-series input matrix The Middle The feature vectors at each time step are denoted as follows: , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The standardized physical augmentation feature vectors at each time step have a dimension of . .

[0103] 2) Establish a device type embedded table Device type embedded table It is a trainable parameter matrix with dimension . ,in, This indicates the number of device types, which can be taken as follows in this invention. These correspond to tower cranes and construction hoists, respectively. This represents the dimension of the device type embedding vector, used to control the capacity for representing device type information; for example, it can be chosen as... or .

[0104] Furthermore, based on the device type label Embedded table from device type Read device type embedding vector , This refers to the device type label. The continuous vector representation is used to enable the time-series risk classification model to learn the differences in risk response of different construction equipment.

[0105] In one implementation, the reading and fusion of the device type embedding vector is achieved through the following steps: First, create a dimension... Trainable embedding matrix ,in, This represents the total number of equipment types. It is the dimension of the embedding vector; then, based on the input device type label... (An integer index) from the matrix The corresponding row vector is retrieved from the list; this vector is the device type embedding vector. This process is equivalent to a table lookup operation. The contents of this table (i.e., the embedding vectors) are continuously optimized through backpropagation during model training, enabling the model to learn differentiated representations for different device types. Finally, at each time step... The normalized physical augmentation feature vector of this time step Device type embedding vector obtained from table lookup By connecting the first and last features along the feature dimension, a device fusion feature vector is formed.

[0106] In one embodiment, for example, the number of device types (Tower cranes, construction hoists), embedded vector dimensions initialize a Embedding matrix For the first A sample, if its device type label (Corresponding to the tower crane), then from Take the 0th row vector as ,For example If the standardized physical augmentation eigenvector at a certain time step Dimensions Its value is The resulting device fusion feature vector after splicing is: Its dimensions are 16.

[0107] 3) At each time step, normalize the physical augmentation feature vector. Embedded vector with device type Concatenating along the feature dimensions yields the device fusion feature vector; since the device type label is a window-level attribute, the device type embedding vector... The process is repeated at each time step within the same time window. Through this operation, the time-series risk classification model can simultaneously obtain information on the current wind field, load, attitude, foundation condition, and construction equipment type when processing each time step.

[0108] 4) Input the device fusion feature vector into the fusion mapping layer to obtain the fusion input vector. fusion of input vectors It refers to the first The number of wind resistance and overturning monitoring samples was in the first The unified feature vector before entering the overturning drive force gated loop unit at each time step has a dimension of 1. , This indicates the dimension of the fused input vector, for example, it can be taken as... .

[0109] In one implementation, the fusion mapping layer consists of a fully connected layer, layer normalization, and a hyperbolic tangent activation function in sequence. The fully connected layer maps the concatenated features from different sources to a unified latent space, the layer normalization stabilizes the numerical distribution of different channels, and the hyperbolic tangent activation function enhances nonlinear representation and limits extreme inputs. The final output is the fused input vector. .

[0110] In one embodiment, as an example, if the total number of feature channels of the physical enhancement timing input matrix... Device type embedding vector dimension Then the dimension of the device fusion feature vector after each time step is... If the input vector dimensions are fused Then the fusion mapping layer will 3D device fusion feature vector mapping is 3D fusion input vector.

[0111] It should be noted that the equipment type embedding is not a separate additional information for the final classification layer, but participates in the fusion of physical enhancement features at each time step, thereby enabling the temporal risk classification model to adaptively understand the risk meaning of the same physical enhancement feature on different construction equipment based on the type of construction equipment.

[0112] S302, Overturning Drive Force Gated Cyclic Unit Construction Sudden overturning risks are often manifested by a rapid increase in the equivalent overturning moment and the equivalent overturning moment growth characteristics within a short period of time. If conventional recurrent neural networks excessively retain historical stable states, the current dangerous input may be diluted by historical information.

[0113] This invention transforms the equivalent overturning moment and its growth characteristics into an activation value for the overturning driving force, and injects this activation value into the gating process of the overturning driving force gating loop unit. This makes it easier to write the current dangerous state into high-risk time steps. The specific steps are as follows: 1) From the normalized physical enhancement timing input matrix The standardized channel values ​​corresponding to the standardized equivalent overturning moment and the standardized channel values ​​corresponding to the growth characteristics of the standardized equivalent overturning moment are read from the database. The standardized equivalent overturning moment is denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The equivalent overturning moment at each time step is a standardized value after grouping by equipment type; the growth characteristic of the standardized equivalent overturning moment is denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The equivalent overturning moment growth characteristics at each time step are standardized values ​​after grouping by equipment type.

[0114] 2) Truncate the negative values ​​in the standardized equivalent overturning moment and the standardized equivalent overturning moment growth characteristics to... A positive driving force input is obtained. This processing is used to make the overturning driving force gating process focus on the overturning driving force and its growth status, which is higher than the conventional level of similar construction equipment. When the standardized equivalent overturning moment is lower than the average level of similar construction equipment, a negative value will not trigger hazard writing enhancement; when the standardized equivalent overturning moment growth characteristic is negative, it indicates that the overturning driving force has not continued to increase, and hazard writing enhancement will not be triggered either.

[0115] In one embodiment, the distance is used as a measure of the standardized equivalent overturning moment. and the growth characteristics of standardized equivalent overturning moment Negative value truncation: If at a certain time step If the value is positive, it indicates that the current overturning driving force is higher than the average level of similar equipment. Retaining this value for subsequent calculations may trigger a dangerous write enhancement; if at a certain time step... A negative value indicates that the current overturning driving force is lower than the average level of similar equipment. A cutoff of 0 indicates that the mechanical state at the current moment is relatively safe and will not have an additional impact on the gating process.

[0116] 3) Map the positive driving force input to the overturning driving force activation value. Overturning driving force activation value It refers to the first The number of wind resistance and overturning monitoring samples was in the first The physical hazard drive intensity at each time step is used to adjust the cyclic unit's selection of current hazard inputs and historical states. The calculation method is expressed as follows: ; in, Indicates the equivalent overturning moment activation scaling factor. This indicates the scaling factor activated by the increase in the equivalent overturning moment. Both factors are used to adjust the contribution of the current overturning driving force level and the rate of increase of the overturning driving force; for example, both can be set to... ; This represents the hyperbolic tangent function, used to limit the activation value of the overturning driving force within a stable range. The larger the activation value of the overturning driving force, the more likely the current time step is to be a moment of increased overturning risk.

[0117] 4) Fuse the input vector The hidden state vector of the previous time step and overturning driving force activation value The overturning driving force gated loop unit is used as a common input to obtain the hidden state vector at the current time step. Among them, the hidden state vector of the previous time step This refers to the period up to the [number]th The risk dynamic memory vector at each time step has a dimension of [missing information]. , Represents the dimension of hidden states, such as the available state. ;when At that time, the initial hidden state vector It can be set to an all-zero vector or a trainable initial vector.

[0118] In its implementation, the internal calculation of the overturning drive force gating loop unit is first based on the fused input vector. and the hidden state vector of the previous time step The basic reset gate vector is calculated according to the mechanism of the standard gated loop unit. Preserve the gate vector with the basic state Then, the overturning driving force is used to activate the value. The two basic gate vectors are modulated, specifically as follows: the final reset gate vector used for subsequent calculations. Reset gate vector by basic and The weighted sum is then passed through the Sigmoid function to obtain the result, making Follow The size of the state is increased, allowing candidate states to reference historical information more comprehensively; ultimately, it is used as the state-preserving gate vector for subsequent calculations. Then the gate vector is preserved by the basic state. and The weighted sum is then passed through the Sigmoid function to obtain the result, making Follow It increases and decreases. Due to this ebb and flow, the activation value of the overturning driving force... As the value increases, the model tends to update the total hidden state using the candidate state at the current time step, thus quickly writing in the current strong danger information. Here, the reset gate vector is denoted as... , Used to control the proportion of the hidden state vector from the previous time step participating in the candidate hidden state calculation; the state preservation gate vector is denoted as... , Used to control the proportion of the hidden state vector from the previous time step retained in the current hidden state vector. Overturning driving force activation value. The reset gate vector is enhanced in a positive manner, while the state retention gate vector is reduced in a negative manner, making it easier to write the current hazard information into the high equivalent overturning moment time step.

[0119] In one embodiment, as an example, suppose that at a certain moment, due to the simultaneous detection of strong gusts of wind and sudden changes in equipment tilt, the calculated activation value of the overturning driving force is... Approaching 1.0. At this point, the modulated reset gate vector... It will tend towards 1, incorporating more past state information into the calculation of the candidate hidden state vector; at the same time, the modulated state-preserving gate vector It will tend towards 0. The hidden state vector at the current time step is updated at the last update. At that time, due to Approaching 0, the hidden state of the previous time step The proportion that is retained will be very small, while the current candidate hidden state vectors containing dangerous information The proportion will be very large, thus enabling the rapid writing of dangerous information.

[0120] 5) Based on the reset gate vector Filter the hidden state vector from the previous time step and combine it with the current fused input vector. Calculate the candidate hidden state vector Candidate hidden state vector This refers to a candidate representation of a risk state formed by the current physical augmentation input and historical information after gating and filtering, with a dimension of [missing information]. .

[0121] In the specific implementation, when calculating the candidate hidden state vector At that time, first set the hidden state vector of the previous time step. With Reset Gate Vector Element-wise multiplication is used to filter historical states; then the filtered historical states are combined with the current fused input vector. The inputs are combined into a fully connected layer and passed through a hyperbolic tangent activation function to obtain candidate hidden state vectors. For example, suppose the gate vector is reset at some point. If some elements in the array are close to 0, then The corresponding historical information will be "reset," and the model will primarily be based on the current input. This generates candidate states to capture sudden dangerous patterns.

[0122] It should be noted that in the calculation of candidate hidden states, the larger the reset gate vector, the higher the proportion of the hidden state vector from the previous time step involved in the calculation; the smaller the reset gate vector, the more the candidate hidden state vector emphasizes the current input information.

[0123] 6) Based on the state-preserving gate vector The hidden state vector of the previous time step With candidate hidden state vector Perform weighted fusion to obtain the hidden state vector at the current time step. The hidden state vector at the current time step It refers to the first The number of wind resistance and overturning monitoring samples was in the first The risk dynamic memory vector after the overturning drive force gating update is completed at each time step, with dimension [missing information]. .

[0124] In the specific implementation, the gate vector is maintained according to the state. When performing weighted fusion, the calculation method is as follows: current hidden state vector equal Element-wise multiplication of the hidden state vector from the previous time step In addition Element-wise multiplication by the candidate hidden state vector This process is a linear interpolation process, which controls the ratio of old to new information retained.

[0125] It should be noted that when the state retention gate vector is large, the hidden state vector at the current time step retains more historical information; when the state retention gate vector is small, the hidden state vector at the current time step is written more into the candidate hidden state vector. Since an increase in the overturning driving force activation value will reduce the state retention gate vector, when strong gusts, strong upwind projection, heavy load tilting, or insufficient foundation bearing capacity occur simultaneously, the current dangerous input can enter the risk dynamic memory vector more quickly.

[0126] In one embodiment, for example, when a sudden gust of wind speed increases at a certain time step, and the wind direction angle and the boom rotation angle form a large wind direction attitude projection factor, the equivalent overturning moment increases. If simultaneously the load weight is large, the equipment tilt angle increases, or the foundation bearing capacity monitoring value is lower than the nominal reference value of the foundation bearing capacity, the equivalent overturning moment increases further, and the overturning driving force activation value increases accordingly. At this time, the overturning driving force gating cycle unit reduces the maintenance of historical stable states and enhances the ability to write current dangerous states, thereby improving the response speed to sudden overturning risks.

[0127] It should also be noted that the overturning drive force gating recurrent unit does not arbitrarily learn the risk time dependency using a black-box recurrent network. Instead, it uses the equivalent overturning moment and its growth characteristics as explicit physical gating signals to directly control the temporal state update process. This design can improve the stability of risk identification under conditions of strong gusts, sudden attitude changes, load eccentricity, and weakened foundation.

[0128] S303, Risk Evidence Attention Pooling and Risk Category Output Evidence of wind overturning risk may appear in the middle of the time window or at the end of the time window. If only the hidden state vector of the last time step is used as the window-level representation, the impact of short-term strong gusts that occur in the middle time step may be missed.

[0129] This invention calculates risk evidence weights based on the hidden state vector and the overturning driving force activation value, and performs weighted aggregation of the hidden state vectors at all time steps to obtain a window-level risk representation. The specific steps are as follows: 1) Read the first The hidden state vector sequence of each wind-resistant overturning monitoring sample at all time steps and overturning driving force activation value sequence Among them, the hidden state vector sequence is used to characterize the risk dynamic information accumulated at each time step, and the overturning driving force activation value sequence is used to characterize the physical hazard driving intensity at each time step.

[0130] 2) Calculate the risk evidence score for each time step. The risk evidence score is denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The contribution of each time step to the window-level risk assessment. In a specific implementation, the hidden state vector can be input into a trainable linear scoring layer, and a weighted term of the overturning driving force activation value can be superimposed to obtain the risk evidence score; the larger the overturning driving force activation value, the easier it is to increase the risk evidence score.

[0131] In its implementation, the trainable linear scoring layer consists of a fully connected layer, which divides the score into four dimensions. The hidden state vector is mapped to a scalar, which is the initial risk evidence score; then, this scalar is scaled by a trainable scaling factor. Weighted overturning drive activation value Add them together to get the final risk evidence score. .

[0132] 3) Calculate the risk evidence weights by applying a normalized exponential function (i.e., the Softmax function) to the risk evidence scores for all time steps. The risk evidence weights are denoted as... , It refers to the first The number of wind resistance and overturning monitoring samples was in the first The weights of each time step participating in window-level risk convergence, and the sum of the risk evidence weights for all time steps, are: The higher the risk evidence weight, the greater the contribution of that time step to the final risk category judgment.

[0133] 4) The hidden state vector sequence is weighted and summed according to the risk evidence weights to obtain the attention risk representation vector. Attention risk representation vector It refers to a window-level risk feature formed by the weighted aggregation of risk evidence from all time steps, used to characterize the dynamic information with the greatest risk discriminative significance within the time window.

[0134] Furthermore, max pooling is performed on the hidden state vector sequence along the time dimension to obtain the peak risk representation vector. Peak risk characterization vector This is used to preserve evidence of short-term, sudden risks. Then, the attention risk representation vector is... With peak risk representation vector By concatenating the vectors, a window-level risk representation vector is obtained. Window-level risk representation vector It refers to the first The final classification input vector is formed by gating the overturning driving force and aggregating the risk evidence from each wind-resistant overturning monitoring sample.

[0135] 5) Window-level risk representation vector Inputting the fully connected classification layer yields the risk category probability distribution vector. Risk category probability distribution vector It refers to the first The predicted probability vectors for each wind-resistant overturning monitoring sample belonging to each risk category have the following dimensions: The three elements correspond to no risk, low risk, and high risk, respectively.

[0136] In practical implementation, the fully connected layer will have a dimension of Window-level risk representation vector Through a linear transformation, the data is mapped to a 3-dimensional vector. The three elements of this vector correspond to the original classification scores for no risk, low risk, and high risk, respectively. Then, these three classification scores are used as input and normalized using the Softmax function, outputting a 3-dimensional probability vector that sums to 1, representing the risk category probability distribution vector. .

[0137] In one embodiment, as an example, if the risk category probability distribution vector Then the maximum prediction probability is Corresponding to the third risk category, the time-series risk classification model outputs a high-risk category; if the risk category probability distribution vector If the highest predicted probability corresponds to the risk-free category, then the time series risk classification model outputs the risk-free category.

[0138] It should be noted that the risk evidence attention pooling does not simply use the hidden state vector at the end of the time window, but integrates the hidden state vectors of all time steps and uses the overturning driving force activation value to guide the model to focus on the time steps with strong physical hazard driving force, thereby improving the ability to identify short-term strong gust impacts and sudden risk changes in the middle of the time.

[0139] S304. Setting the Category-Weighted Loss Function and Model Training The wind overturning monitoring data at construction sites often suffers from class imbalance, with a large number of risk-free samples and a small number of high-risk samples. If ordinary cross-entropy loss is used directly, the time series risk classification model is prone to bias towards the risk-free category.

[0140] This invention uses category-weighted cross-entropy loss to train a time-series risk classification model and uses macro-average F1 score to select the final model parameters. The specific steps are as follows: 1) The first Risk category label for individual wind resistance and overturning monitoring samples Encoded as one-hot tag vectors One-hot tag vector It refers to the first The risk category monitoring signal corresponding to each wind resistance overturning monitoring sample is dimensional as follows: .

[0141] In practical implementation, if the first If a wind-resistant overturning monitoring sample belongs to the no-risk category, then the unique thermal label vector is: If it belongs to the low-risk category, then the unique hot tag vector is... If it belongs to a high-risk category, then the unique hot tag vector is... .

[0142] 2) Calculate the class weights based on the number of samples in each risk category in the training sample set. Specifically, the first... The number of training samples for each risk category is denoted as . , This represents the risk category index, with values... The number of risk category samples with the largest number in the training sample set is denoted as . ;No. The category weights of each risk category are denoted as: You can press Calculate, where, This represents the stability constant for calculating class weights, used to prevent the denominator from being too large. For example, it is advisable Risk categories with fewer samples have higher weights, thus receiving more attention during training.

[0143] 3) Calculate the category-weighted cross-entropy loss for the wind resistance overturning monitoring samples in a single training batch.

[0144] In practical implementation, the number of training batch samples is denoted as... , This indicates the number of wind-resistant overturning monitoring samples used in a single parameter update, for example, it can be taken as... or The category-weighted cross-entropy loss is calculated based on the predicted probability and category weight corresponding to the true risk category. When high-risk category samples are misclassified, the loss increases more significantly due to the larger weight of the high-risk category, thus prompting the time-series risk classification model to improve its ability to identify high-risk categories.

[0145] 4) Update the trainable parameters of the time-series risk classification model using error backpropagation. In one implementation, the Adam optimizer can be used for training, with an initial learning rate of [value missing]. The number of training cycles can be taken as follows: to When the validation set loss is continuous If the learning rate does not decrease over a training cycle, multiply it by 1. Training can be stopped early if the average F1 score of the validation set macro does not improve for several consecutive training cycles.

[0146] 5) After each training cycle, evaluate the time-series risk classification model using the validation sample set. Specifically, calculate the F1 score for each of the three risk categories: no risk, low risk, and high risk, and take the arithmetic mean of the three F1 scores to obtain the macro-average F1 score. Save the model parameters corresponding to the highest macro-average F1 score as the final time-series risk classification model parameters.

[0147] S4. Online wind-resistant overturning risk assessment of construction equipment After training the time-series risk classification model, the final model is deployed to the online monitoring system for construction equipment. During online operation, the system continuously collects multi-source monitoring data from construction equipment and, following the same data processing flow as the training phase, forms a standardized physical augmentation time-series input matrix. It then outputs the current risk level and alarm results. The specific steps are as follows: 1) Extract the wind resistance overturning monitoring sample to be evaluated from the online monitoring data stream using a sliding window. The wind resistance overturning monitoring sample to be evaluated is denoted as... , This refers to a segment of continuous monitoring data used for wind-resistant overturning risk assessment at the current moment. The sliding window length remains consistent with the training phase; for example, taking... seconds; sliding step size can be selected. Second, seconds or Seconds. The smaller the sliding step size, the more frequently the risk assessment is updated; the larger the sliding step size, the lower the computational burden.

[0148] 2) The same physical guidance feature reconstruction and enhancement processing as the training phase was performed on the wind-resistant overturning monitoring samples to be evaluated. Specifically, time alignment, outlier pruning, and construction of the basic monitoring time series matrix were first performed; then, the coupling characteristics of gust tilt abrupt changes, equivalent overturning moment, equivalent overturning moment growth characteristics, and foundation softening cumulative index were calculated; finally, based on the equipment type labels of the wind-resistant overturning monitoring samples to be evaluated, the equipment type grouping statistics saved in the training phase were selected to obtain the standardized physical enhancement time series input matrix to be evaluated. The normalized physical enhancement timing input matrix to be evaluated It refers to the model input matrix after the wind resistance overturning monitoring samples to be evaluated have undergone physical guidance feature reconstruction, enhancement, and equipment type grouping standardization.

[0149] It should be noted that during the online inference phase, the mean, standard deviation, equipment tilt angle difference mean, or equipment tilt angle difference standard deviation of the characteristic channels must not be recalculated based on the wind-resistant overturning monitoring samples to be evaluated. Instead, the statistics saved during the training phase must be used. Otherwise, the risk magnitude within the wind-resistant overturning monitoring samples to be evaluated may be rescaled, leading to unstable risk assessment results.

[0150] 3) Input the normalized physical enhancement timing matrix to be evaluated. The corresponding equipment type label is input into the final temporal risk classification model. The final temporal risk classification model sequentially performs equipment type embedding and physical enhancement feature fusion mapping, overturning drive force gated recurrent encoding, risk evidence attention pooling, and risk category output to obtain the probability distribution vector of the risk category to be evaluated. The probability distribution vector of the risk category to be evaluated It refers to the predicted probability that the wind-resistant overturning monitoring sample to be evaluated belongs to one of the three risk categories: no risk, low risk, and high risk.

[0151] 4) Based on the probability distribution vector of the risk category to be assessed Output the current risk level.

[0152] In practical implementation, the confidence threshold is denoted as... , This refers to the minimum forecast probability required to accept the current predicted risk level, for example, the minimum probability that can be obtained. to If the maximum predicted probability in the probability distribution vector of the risk category to be evaluated is less than the confidence threshold... If the probability is not less than the confidence threshold, a low confidence warning will be output, and the subsequent sliding window will be observed; if the maximum predicted probability is not less than the confidence threshold, the output will be positive. If the maximum predicted probability is obtained, then the risk category corresponding to that probability will be output.

[0153] In one embodiment, as an example, if the probability distribution vector of the risk category to be evaluated is... And confidence threshold Then the maximum prediction probability is For high-risk categories, the online monitoring system outputs high-risk results; if the probability distribution vector of the risk category to be assessed is... The maximum predicted probability is Below the confidence threshold If the online monitoring system outputs a low confidence level alert, it will wait for further confirmation via a subsequent sliding window.

[0154] 5) Perform time consistency verification on risk levels for multiple consecutive sliding windows. The number of consecutive verification windows is denoted as... This corresponds to the number of sliding windows that need to continuously meet the same risk condition before triggering a formal alarm. For example, it can be taken as... to When the high-risk category is in continuous Each sliding window contains the category with the highest predicted probability, and the corresponding predicted probability is not less than the confidence threshold. When the risk level is high, a formal high-risk alarm is output; when the low-risk category is continuously high, a formal high-risk alarm is output. When all sliding windows contain the category with the highest predicted probability, a low-risk warning is output; when the risk categories of adjacent sliding windows change frequently or the predicted probabilities are low, a formal alarm is not triggered, and subsequent monitoring data continues to be collected.

[0155] 6) Along with the risk level, supplementary explanatory information is also output. This information includes the maximum gust tilt change coupling characteristic, maximum equivalent overturning moment, maximum equivalent overturning moment growth characteristic, and foundation softening accumulation index within the current sliding window. A high maximum gust tilt change coupling characteristic indicates that the current risk is primarily related to the simultaneous occurrence of gusts and attitude changes; a high maximum equivalent overturning moment indicates that the current risk is primarily related to the superposition of wind load, load eccentricity, or insufficient foundation bearing capacity; and a high foundation softening accumulation index indicates that the current risk includes factors contributing to the continuous deterioration of foundation bearing capacity. By outputting this supplementary explanatory information, on-site management personnel can more quickly identify the source of the risk and take appropriate measures such as load reduction, shutdown, adjusting the boom direction, reinforcing the foundation, or evacuating the danger zone.

[0156] It should be noted that the online wind-resistant overturning risk assessment process uses the exact same feature reconstruction logic and standardized statistics as the training phase, ensuring consistent data distribution between training and inference. Time consistency verification reduces false alarms caused by abnormal noise in individual sliding windows; supplementary interpretation information improves the usability of risk assessment results in construction site safety management.

[0157] In one embodiment, such as Figure 2 As shown, the process of constructing coupled features of gust tilt change is analyzed. This experiment consists of three sub-figures: (a), (b), and (c). In (a), the horizontal axis represents the time step (steps), and the vertical axis represents the wind speed (m / s). The instantaneous gust wind speed and its local average gust wind speed are plotted simultaneously, demonstrating that the enhancement degree of the instantaneous gust relative to the local background (gust enhancement ratio) can be quantitatively extracted. In (b), the horizontal axis represents the time step (steps), the left vertical axis represents the equipment tilt angle difference Δφ (degrees / step), and the right vertical axis represents the change intensity (dimensionless). The curve shows that the attitude change suddenly increases at the peak of the gust, and the change intensity increases significantly, reflecting the detectability of attitude change. (c) shows the time step (step) on the horizontal axis and the coupling feature value (dimensionless) on the vertical axis. The comparison shows the product of the gust enhancement ratio and the change intensity without hyperbolic tangent compression and the gust tilt change coupling feature obtained after hyperbolic tangent function compression. It proves that the coupling feature can output a high value when gust enhancement and attitude change occur simultaneously, while maintaining a low value when there are only wind speed fluctuations or normal attitude changes, effectively suppressing false alarms caused by single-channel noise.

[0158] In one embodiment, such as Figure 3As shown, the online wind overturning risk assessment output is analyzed. This experiment consists of two sub-graphs, (a) and (b). In (a), the horizontal axis represents the sliding window number, and the vertical axis represents the probability value predicted by the model (0-1). The green, orange, and red curves represent the probabilities of no risk, low risk, and high risk, respectively, and the gray dashed line represents the confidence threshold η=0.7. In (b), the horizontal axis also represents the sliding window number, and the vertical axis represents the binary flag (0-1). The red stepped line represents the high-risk flag in a single window, and the black stepped line represents the formal high-risk alarm triggered after three consecutive windows of time consistency confirmation. This technology not only outputs the risk probability in each sliding window but also effectively filters single-window noise fluctuations through confidence threshold and time continuity confirmation, producing stable and reliable online alarm results.

[0159] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters, characterized in that, Includes the following steps: S1. Obtain wind-resistant overturning monitoring samples and construct a training sample set containing equipment type labels, risk category labels, and data source labels; S2. Through numerical calibration by equipment type grouping, coupling of gust tilt change, reconstruction of equivalent overturning moment, cumulative modeling of foundation softening, and splicing of physical enhancement features, the original monitoring quantities are transformed into a physical enhancement time series input matrix with clear mechanical meaning. The physical enhancement time series input matrix is ​​then standardized channel by channel to obtain a standardized physical enhancement time series input matrix. S3. Construct an adaptive temporal risk classification model for equipment. The adaptive temporal risk classification model for equipment is trained by embedding equipment type, fusing physical enhancement features, gating recurrent units for overturning driving force, attention pooling of risk evidence, and category-weighted loss to achieve adaptive risk identification of equipment. S4. Continuously collect multi-source monitoring data of construction equipment, construct a standardized physical augmentation time-series input matrix, and input the standardized physical augmentation time-series input matrix into the trained equipment adaptive time-series risk classification model to obtain the risk level.

2. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 1, characterized in that, In step S1, the first Each time window serves as a sample for wind resistance and overturning monitoring. , This represents the sample index, with a value range of [value range missing]. , This represents the total number of wind resistance overturning monitoring samples, and each wind resistance overturning monitoring sample contains... Each time step Indicates the number of time steps within a single time window; the first Multi-source monitoring data for each wind resistance overturning monitoring sample includes continuous wind speed. Instantaneous gust wind speed Wind angle boom rotation angle Lifting weight Equipment tilt angle Foundation bearing capacity monitoring values and equipment operating status , This represents the time step index, with a value range of [value range missing]. Configure equipment type labels for each wind-resistant overturning monitoring sample. and risk category labels Record the data source label for each wind resistance overturning monitoring sample. .

3. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 1, characterized in that, In step S2, data alignment, outlier handling, and grouping statistics are completed for the wind-induced overturning monitoring samples based on the equipment type labels, resulting in the basic monitoring time series matrix: For the first The timestamps of each sensor in the wind-resistant overturning monitoring sample are aligned to ensure that the continuous wind speed, instantaneous gust wind speed, wind direction angle, boom rotation angle, load weight, equipment tilt angle, foundation bearing capacity monitoring values, and equipment operating status correspond at the same time step; outliers exceeding the physical range of the sensors are clipped; and the original monitoring data after time alignment and outlier processing are spliced ​​together in a fixed channel order to obtain the basic monitoring time series matrix. Within the training sample set, the device type labels are statistically analyzed according to the device type labels. Feature channel mean and characteristic channel standard deviation .

4. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 1, characterized in that, In step S2, the enhancement of instantaneous gust speed relative to the local gust background is coupled with the abrupt change intensity of the equipment tilt angle relative to the normal fluctuation scale of similar construction equipment to obtain the gust tilt abrupt change coupling characteristics: From the Instantaneous gust wind speed was read from a sample of wind resistance and overturning monitoring data. and equipment tilt angle Calculate the difference in device tilt angle between adjacent time steps. Based on the equipment type labels, the equipment tilt angle difference values ​​are grouped and statistically analyzed in the training set to obtain the equipment type labels. Mean difference in equipment tilt angle Standard deviation of equipment tilt angle difference ; Using the current time step as the endpoint, calculate the local average value of the instantaneous gust wind speed, and denot the local average gust wind speed as follows: ; The instantaneous gust speed at the current time step Local average gust wind speed The ratio was calculated to obtain the gust enhancement ratio. According to the first Equipment type label for individual wind resistance overturning monitoring samples Select the average difference of the corresponding equipment tilt angle. Standard deviation of equipment tilt angle difference Calculate the intensity of sudden changes in the tilt angle of the equipment. ; Increase the ratio of gusts and the intensity of sudden changes in equipment tilt angle By performing product coupling and compressing extreme values ​​using a hyperbolic tangent function, the coupling characteristics of gust tilt abrupt changes are obtained. .

5. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 1, characterized in that, In step S2, the multi-source monitoring data is uniformly transformed into driving force characteristics for overturning risk. Based on the overturning force rectangularity mechanism, the equivalent overturning moment is reconstructed, thereby obtaining the equivalent overturning moment growth characteristics: Read the first The number of wind resistance and overturning monitoring samples was in the first Continuous wind speed at each time step and instantaneous gust wind speed The maximum of the two values ​​is taken as the effective wind speed. ; Read the first The number of wind resistance and overturning monitoring samples was in the first Wind angle at each time step and boom rotation angle Calculate the relative angle difference between the two, and normalize the relative angle difference to... The effective windward projection degree is calculated within the range to obtain the wind direction and attitude projection factor. According to the effective wind speed Wind direction and attitude projection factor Comprehensive coefficient of pneumatic boom corresponding to equipment type Calculate the wind load moment term The comprehensive coefficient of the pneumatic boom refers to the equipment type label. The wind load calibration factor; Read the first The number of wind resistance and overturning monitoring samples was in the first Lifting weight at each time step and equipment tilt angle Calculate the eccentric moment term of the hoisting load. ; Read the first The number of wind resistance and overturning monitoring samples was in the first Foundation bearing capacity monitoring values ​​at each time step and the nominal reference value of foundation bearing capacity By comparison, the calculated foundation weakening moment term is denoted as... ; Wind load moment term 1. Lifting eccentricity moment item and foundation weakening moment term Adding them together, we get the equivalent overturning moment. Based on the equivalent overturning moment, the growth characteristics of the equivalent overturning moment were calculated. , , Indicates the first The number of wind resistance and overturning monitoring samples was in the first The equivalent overturning moment at each time step.

6. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 1, characterized in that, In step S2, based on the extent of foundation bearing capacity loss within the time window, the proportion of time with insufficient foundation bearing capacity, and the longest continuous insufficient time, a foundation softening cumulative index is constructed: For the first For each time step of the wind-resistant overturning monitoring sample, the extent of foundation bearing capacity loss is calculated. The extent of the foundation bearing capacity loss is the nominal reference value of the foundation bearing capacity. Higher than the monitoring value of foundation bearing capacity Part; Statistics The number of time steps within the time window where the foundation bearing capacity of the wind-resistant overturning monitoring sample is insufficient. The number of time steps for insufficient foundation bearing capacity Divide by the time window length The percentage of time with insufficient foundation bearing capacity was obtained. The number of time steps where the foundation bearing capacity is insufficient is the total number of time steps in which the monitored foundation bearing capacity value is less than the nominal reference value of the foundation bearing capacity. Statistics The longest consecutive insufficient time step within the time window of each wind-resistant overturning monitoring sample. The longest continuous insufficient time step number is the longest time step during which the foundation bearing capacity monitoring value is continuously lower than the nominal reference value of the foundation bearing capacity. The average bearing capacity loss of the foundation within the calculation time window. The average bearing capacity deficit, the percentage of time with insufficient bearing capacity, and the number of longest consecutive insufficient time steps are combined to obtain the cumulative index of foundation softening. The formula is expressed as: ,in, Indicates the continuous deficiency penalty coefficient; and the cumulative index of foundation softening. Copy to the Each time step of the wind-resistant overturning monitoring sample forms a time series channel for the cumulative index of foundation softening.

7. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 1, characterized in that, In step S3, the device type embedding and physical enhancement feature fusion mapping are performed: Read the first Standardized physical augmentation time-series input matrix for one wind resistance overturning monitoring sample and device type label ; Create a device type embedded table Device type embedded table For trainable parameter matrices, based on device type labels Embedded table from device type Read device type embedding vector At each time step, the standardized physical augmentation feature vector is... Embedded vector with device type The device fusion feature vector is obtained by concatenating along the feature dimension; the device fusion feature vector is then input into the fusion mapping layer to obtain the fusion input vector. The fusion mapping layer consists of a fully connected layer, a layer normalization layer, and a hyperbolic tangent activation function in sequence.

8. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 7, characterized in that, In step S3, the overturning drive force gating cycle unit specifically includes: From the normalized physical enhancement time-series input matrix Read the standardized channel values ​​corresponding to the equivalent overturning moment and the standardized channel values ​​corresponding to the growth characteristics of the equivalent overturning moment; truncate the negative values ​​in the standardized equivalent overturning moment and the standardized equivalent overturning moment growth characteristics. The positive driving force input is obtained, and the positive driving force input is mapped to the overturning driving force activation value. ; fuse the input vector The hidden state vector of the previous time step and overturning driving force activation value The overturning driving force gated loop unit is used as a common input to obtain the hidden state vector at the current time step. The specific process is as follows: The overturning drive force gated loop unit is based on a fused input vector. and the hidden state vector of the previous time step The basic reset gate vector is calculated according to the mechanism of the standard gated loop unit. Preserve the gate vector with the basic state Activation value using overturning driving force Modulating these two basic gate vectors yields the reset gate vector. and state-preserving gate vector ; Based on the reset gate vector Filter the hidden state vector from the previous time step and combine it with the current fused input vector. Calculate the candidate hidden state vector : The hidden state vector of the previous time step With Reset Gate Vector Element-wise multiplication is used to filter historical states, and the filtered historical states are then fused with the current input vector. The inputs are combined into a fully connected layer and passed through a hyperbolic tangent activation function to obtain candidate hidden state vectors. ; Based on the state preservation gate vector The hidden state vector of the previous time step With candidate hidden state vector Perform weighted fusion to obtain the hidden state vector at the current time step. .

9. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 7, characterized in that, In step S3, the risk evidence attention pooling and risk category output specifically include: Read the first The hidden state vector sequence of each wind-resistant overturning monitoring sample at all time steps and overturning driving force activation value sequence Calculate the risk evidence score for each time step. The hidden state vector is input into a trainable linear scoring layer, and a weighted term of the overturning driving force activation value is superimposed to obtain the risk evidence score; the trainable linear scoring layer consists of a fully connected layer. The risk evidence weights are calculated by applying a normalized exponential function to the risk evidence scores for all time steps. The attention risk representation vector is obtained by weighting and summing the hidden state vector sequence according to the risk evidence weights. Max pooling is performed on the hidden state vector sequence along the time dimension to obtain the peak risk representation vector. , and the attention risk representation vector With peak risk representation vector By concatenating the vectors, a window-level risk representation vector is obtained. ; Window-level risk representation vector Inputting the fully connected classification layer yields the risk category probability distribution vector. The risk category corresponding to the highest predicted probability is output as the final risk category.

10. The method for predicting the wind and overturning risk of on-site equipment by combining environmental parameters according to claim 9, characterized in that, In step S3, the device adaptive time-series risk classification model is trained using category-weighted cross-entropy loss, and the final model parameters are selected using macro-average F1 score.

Images