Crane AI monitoring method and system considering wind load

Abstract

The invention relates to the technical field of crane safety monitoring, in particular to a crane AI monitoring method and system considering wind load. Wind speed, wind direction and strain data are collected through a sensor, discrete sections of a tower body are divided, a wind pressure distribution map is calculated, and a wind load strain component is extracted from the strain data. And inputting the atlas and the components into a deep neural network, evaluating the similarity between the current state and the instability mode, and predicting a critical moment to calculate a safety margin. According to the method, accurate prediction and active early warning of the crane instability risk under the wind load are realized.

Description

Technical Field

[0001] This invention relates to the field of crane safety monitoring technology, and in particular to an AI monitoring method and system for cranes that takes wind load into account. Background Technology

[0002] In the field of crane safety monitoring, existing technologies typically rely on direct measurement of structural stress or deformation to assess its operational status. The conventional approach involves installing a limited number of sensors at critical crane components, such as the tower base or boom, to collect strain or vibration data. This data is then compared in real-time with preset safety thresholds; if a monitored value exceeds the threshold, the system triggers an alarm or initiates a shutdown. The core of this method lies in treating the crane structure as a whole and assuming a direct, easily thresholded correspondence between its load state and structural response.

[0003] However, this conventional monitoring method has significant limitations. In actual operation, especially at heights outdoors, cranes experience dynamic and spatially uneven wind loads. The force of the wind varies with height, producing differentiated effects at different parts of the tower and boom. Existing methods, using discrete point measurements and overall threshold judgments, struggle to accurately reflect this complex and non-uniform wind load effect. The system cannot effectively distinguish between strain caused by the weight of the lifted cargo and strain components caused by complex wind fields, leading to inaccurate assessments of the structure's true stress state.

[0004] This deficiency directly impacts the reliability and timeliness of monitoring systems. Because the spatial distribution of wind loads and their dynamic effects on various structural components cannot be precisely quantified, the system may be slow to respond to potential risks, failing to provide sufficient warning time before wind-induced instability occurs. Conversely, the inability to effectively isolate the effects of wind loads may lead to misjudging structural responses under normal wind loads as dangerous conditions, resulting in frequent false alarms and inappropriate operational interruptions, severely impacting the efficiency and economy of crane operations. Therefore, there is an urgent need for a monitoring technology that can more accurately characterize the spatial distribution and dynamic evolution of wind loads, and based on this, conduct advanced and reliable stability assessments. Summary of the Invention

[0005] The present invention provides a crane AI monitoring method and system that takes wind load into account, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides an AI monitoring method for cranes that takes wind load into account, comprising:

[0007] Wind speed and wind direction data are collected by wind speed sensors and wind direction sensors placed at different heights on the crane tower, and strain data are collected by strain sensors placed on the boom.

[0008] The tower is divided into multiple discrete segments according to its height. The local wind speed value is calculated based on the wind speed data of each discrete segment. The wind pressure of each discrete segment is calculated by combining the windward area and wind resistance coefficient. The segments are arranged according to their spatial location to form a wind pressure distribution map.

[0009] Obtain strain calibration data under no-load and full-load conditions, represent the strain data as a weighted combination of no-load strain calibration data and full-load strain calibration data, solve the weight coefficient by minimizing the fitting error, and extract the wind load strain component from the strain data based on the weight coefficient.

[0010] The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network, and the similarity between the current state and the instability mode is output based on historical instability accident samples.

[0011] The evolution curve is obtained by nonlinear fitting of the numerical change of similarity over time. The critical moment when the similarity reaches the instability threshold is determined based on the evolution curve.

[0012] The safety margin is calculated based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, an early warning signal and action restriction command are generated.

[0013] The tower is divided into multiple discrete segments according to its height. Local wind speed values ​​are calculated based on wind speed data for each discrete segment. Wind pressure in each discrete segment is calculated by combining the windward area and drag coefficient. These segments are then arranged spatially to form a wind pressure distribution map, including:

[0014] Based on the tower structure parameters, the tower is divided into multiple discrete segments along the height direction at preset intervals, and the center height coordinates and windward area of ​​each discrete segment are recorded.

[0015] Read the wind speed data output by the wind speed sensor at the corresponding height position of each discrete segment as the local wind speed value, and read the wind direction data output by the wind direction sensor;

[0016] Calculate the difference in local wind speed between adjacent discrete segments. When the difference exceeds a preset gradient threshold, insert a transition segment between the two discrete segments, perform linear interpolation on the wind speed value of the transition segment, and calculate the wind pressure of the transition segment.

[0017] The bending moment contribution component is obtained by projecting the wind pressure of each discrete segment onto the central axis of the tower. The bending moment contribution components of all discrete segments are accumulated along the height direction to obtain the cumulative bending moment distribution curve. The position with the maximum gradient in the cumulative bending moment distribution curve is identified as the stress concentration height.

[0018] The center height coordinates, wind pressure, and bending moment contribution components of each discrete segment are constructed into a three-dimensional data array. The discrete segments corresponding to the stress concentration height are marked and arranged in ascending order of center height coordinates to form a wind pressure distribution map containing stress concentration location information.

[0019] Obtain strain calibration data under no-load and full-load conditions, represent the strain data as a weighted combination of no-load and full-load strain calibration data, solve for the weighting coefficients by minimizing the fitting error, and extract the wind load strain components from the strain data based on the weighting coefficients, including:

[0020] The crane is controlled to be in no-load and full-load states in a windless environment. The output of the strain sensor is collected and marked as no-load strain calibration data and full-load strain calibration data.

[0021] Strain data of different combinations of lifting weight and wind speed under actual working conditions were collected. A three-dimensional response surface of strain data, lifting weight and wind speed was established. The strain value corresponding to zero lifting weight on the three-dimensional response surface was extracted as the pure wind load strain benchmark, and the strain value corresponding to zero wind speed was extracted as the pure lifting weight strain benchmark.

[0022] The correlation coefficient between the pure suspended load strain benchmark and the no-load strain calibration data and the full-load strain calibration data is calculated as the initial weighting coefficient;

[0023] A search space is constructed with the initial weight coefficients as the center. The weight coefficients in the search space are traversed. The unloaded strain calibration data and the full-load strain calibration data are weighted and combined according to each weight coefficient to obtain the reconstructed strain. The fitting error between the reconstructed strain and the pure suspended strain benchmark is calculated. The weight coefficient with the smallest fitting error is selected as the optimal weight coefficient.

[0024] The wind load strain components are obtained by weighting the unloaded strain calibration data and the full-load strain calibration data according to the optimal weighting coefficient, and then subtracting the weighted combination result from the strain data.

[0025] Strain data were collected under actual operating conditions with different combinations of lifting weight and wind speed. A three-dimensional response surface was established based on the strain data and the relationship between lifting weight and wind speed, including:

[0026] Set the range of load variation and wind speed variation. Within the range of load variation, select multiple load sampling points evenly, and within the range of wind speed variation, select multiple wind speed sampling points evenly. Combine the load sampling points and wind speed sampling points by Cartesian product to obtain the set of sampling working condition points.

[0027] The crane is controlled to execute the lifting weight and wind speed conditions corresponding to each sampling point in the sampling point set one by one, and the output values ​​of the strain sensor corresponding to each sampling point are collected to establish a mapping relationship between the sampling points and the output values ​​of the strain sensor.

[0028] The sampling points in the sampling point set that have missing strain sensor output values ​​are marked as missing points. The neighboring points around the missing points that have collected strain sensor output values ​​are extracted. The interpolated strain value of the missing point is calculated by weighted interpolation based on the strain sensor output values ​​of the neighboring points and the distance between the neighboring points and the missing point.

[0029] The load value of the sampling point is used as the horizontal axis coordinate, the wind speed value of the sampling point is used as the vertical axis coordinate, and the output value of the strain sensor and the interpolated strain value corresponding to the sampling point are used as the vertical axis coordinate to construct a three-dimensional coordinate point cloud.

[0030] A surface fitting operation is performed on a 3D coordinate point cloud to generate a 3D response surface.

[0031] The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network. Based on historical instability accident samples, the similarity between the current state and the instability mode is output, including:

[0032] The wind pressure distribution map and the wind load strain components are time-aligned to construct a nonlinear coupled response function containing time lag terms, thus obtaining the wind pressure-strain coupled response field. The coupled evolution trajectory sequence is obtained by sliding decomposition of the wind pressure-strain coupled response field along the time dimension.

[0033] The coupled evolution trajectory sequence is input into the spatial coding subnetwork and temporal coding subnetwork of the deep neural network to extract spatial gradient features and evolution trend features, respectively. The spatial gradient features and evolution trend features are then input into the physical constraint module. The physical constraint module applies consistency constraints based on the mechanical equilibrium relationship between wind pressure and strain to obtain constraint feature vectors.

[0034] A coupled consistency loss function containing wind pressure-driven error terms and strain inversion error terms is constructed to train a deep neural network, so that the constraint feature vector satisfies the minimum bidirectional prediction error of wind pressure and strain.

[0035] The current coupled evolution trajectory sequence is dynamically time-warped and matched with the coupled evolution trajectory sequence of historical instability accident samples to calculate the trajectory matching distance. The instability mode with the smallest trajectory matching distance is selected as the target instability mode.

[0036] Calculate the path deviation and evolution trend consistency coefficient between the current coupled evolution trajectory sequence and the target instability mode trajectory sequence, and calculate the similarity based on the path deviation and evolution trend consistency coefficient.

[0037] The evolution curve is obtained by nonlinearly fitting the numerical changes of similarity over time. Based on the evolution curve, the critical moments when the similarity reaches the instability threshold are determined, including:

[0038] The similarity scores are collected continuously at multiple times according to a preset sampling period and arranged in chronological order to form a similarity time series;

[0039] The similarity time series is segmented by a sliding window and the rate of change of similarity within each segment is calculated. The time and similarity corresponding to the segments with positive rate of change are extracted to form the unstable trend data.

[0040] The similarity between adjacent sample points in the unstable trend data is subjected to second-order difference operation to obtain a second-order difference value sequence. The absolute value of each second-order difference value in the second-order difference value sequence is normalized and used as the fitting weight coefficient of the corresponding sample point.

[0041] A polynomial function is used to perform weighted curve fitting on the sample points in the unstable trend data to obtain the initial evolution curve and calculate the fitting residual of each sample point. When the fitting residual exceeds the preset residual threshold, the degree of the polynomial function is increased by a preset step size value and the sample points are re-fitted with weighted curve to obtain the optimized evolution curve.

[0042] The instability threshold is substituted into the time variable of the optimization evolution curve solution as the critical time when the similarity reaches the instability threshold.

[0043] The safety margin is calculated based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, an early warning signal and action restriction command are generated, including:

[0044] Calculate the time difference between the critical moment and the current moment;

[0045] Obtain the current lifting mass, amplitude value, and slewing angular velocity of the crane. Input the lifting mass, amplitude value, and slewing angular velocity into the historical instability sample library for matching. Extract the actual braking duration of the target historical sample record with the highest matching degree. Subtract the actual braking duration from the time difference to obtain the controllable time window.

[0046] Obtain the current wind pressure distribution map, input the current wind pressure distribution map into the historical instability sample library for map shape comparison, and count the proportion of samples with wind load abrupt changes in the candidate historical samples with matching shapes as the abrupt change probability. Based on the abrupt change probability, compress the controllable time window to obtain the safety margin.

[0047] When the safety margin is less than the preset safety value, the safety margin is entered into the historical instability sample library for matching, and the slewing speed limit, lifting speed limit and amplitude speed limit recorded in the reference historical sample with the highest matching degree are extracted.

[0048] The safety margin and mutation probability are encapsulated to generate an early warning signal. The slewing speed limit, hoisting speed limit, and amplitude speed limit are encapsulated to generate a motion restriction command. The early warning signal is sent to the monitoring terminal, and the motion restriction command is sent to the crane controller.

[0049] A second aspect of the present invention provides an AI monitoring system for cranes that takes wind load into account, comprising:

[0050] The data acquisition unit is used to collect wind speed and wind direction data through wind speed sensors and wind direction sensors arranged at different height positions on the crane tower, and to collect strain data through strain sensors arranged on the boom.

[0051] The wind pressure calculation unit is used to divide the tower into multiple discrete segments according to its height, calculate the local wind speed value based on the wind speed data of each discrete segment, and calculate the wind pressure of each discrete segment by combining the windward area and wind resistance coefficient, and arrange them according to spatial location to form a wind pressure distribution map.

[0052] The strain analysis unit is used to acquire strain calibration data under no-load and full-load conditions. It represents the strain data as a weighted combination of no-load strain calibration data and full-load strain calibration data. The weight coefficients are solved by minimizing the fitting error, and the wind load strain components are extracted from the strain data based on the weight coefficients.

[0053] The instability prediction unit is used to input the wind pressure distribution map and wind load strain components into a pre-trained deep neural network, and output the similarity between the current state and the instability mode based on historical instability accident samples.

[0054] The critical judgment unit is used to perform nonlinear fitting on the numerical change of similarity over time to obtain the evolution curve, and to determine the critical moment when the similarity reaches the instability threshold based on the evolution curve.

[0055] The early warning generation unit is used to calculate the safety margin based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, it generates an early warning signal and an action restriction command.

[0056] A third aspect of the present invention provides an electronic device, comprising:

[0057] processor;

[0058] Memory used to store processor-executable instructions;

[0059] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0060] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0061] This method divides the tower into multiple discrete segments and calculates the wind pressure in each segment to form a wind pressure distribution map. This accurately reflects the spatial distribution characteristics of wind loads on the crane structure, overcoming the limitations of traditional methods that simplify wind loads to concentrated forces or uniformly distributed loads. Utilizing strain calibration data under no-load and full-load conditions, a weighted combination model separates the wind load strain components from the measured strain data, effectively eliminating the influence of factors such as lifting loads and structural self-weight. The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network, which integrates the structural wind pressure input and structural response output. This allows for in-depth analysis of the intrinsic correlation between the current state and historical instability modes, resulting in a more accurate instability similarity calculation. This makes the risk assessment results more convincing and valuable for early warning. By performing nonlinear fitting on the similarity time series to obtain the evolution curve and predicting the critical moment when the instability threshold is reached, a forward-looking judgment of risk development trends is achieved. Calculating dynamic safety margins based on critical moments and generating early warnings and limiting instructions accordingly transforms post-event alarms into pre-event prevention, providing operators with valuable reaction time and effectively improving the safety and controllability of cranes operating in complex wind farm environments. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating the AI ​​monitoring method for cranes that takes wind load into account in this embodiment.

[0063] Figure 2 This is a flowchart of the three-dimensional response surface generation process in this embodiment. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0065] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0066] Figure 1 This is a flowchart illustrating the AI ​​monitoring method for cranes considering wind loads according to an embodiment of the present invention, as shown below. Figure 1 As shown, the AI ​​monitoring method for cranes considering wind load includes:

[0067] Wind speed and wind direction data are collected by wind speed sensors and wind direction sensors placed at different heights on the crane tower, and strain data are collected by strain sensors placed on the boom.

[0068] The tower is divided into multiple discrete segments according to its height. The local wind speed value is calculated based on the wind speed data of each discrete segment. The wind pressure of each discrete segment is calculated by combining the windward area and wind resistance coefficient. The segments are arranged according to their spatial location to form a wind pressure distribution map.

[0069] Obtain strain calibration data under no-load and full-load conditions, represent the strain data as a weighted combination of no-load strain calibration data and full-load strain calibration data, solve the weight coefficient by minimizing the fitting error, and extract the wind load strain component from the strain data based on the weight coefficient.

[0070] The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network, and the similarity between the current state and the instability mode is output based on historical instability accident samples.

[0071] The evolution curve is obtained by nonlinear fitting of the numerical change of similarity over time. The critical moment when the similarity reaches the instability threshold is determined based on the evolution curve.

[0072] The safety margin is calculated based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, an early warning signal and action restriction command are generated.

[0073] The tower is divided into multiple discrete segments according to its height. Local wind speed values ​​are calculated based on wind speed data for each discrete segment. Wind pressure in each discrete segment is calculated by combining the windward area and drag coefficient. These segments are then arranged spatially to form a wind pressure distribution map, including:

[0074] Based on the tower structure parameters, the tower is divided into multiple discrete segments along the height direction at preset intervals, and the center height coordinates and windward area of ​​each discrete segment are recorded.

[0075] Read the wind speed data output by the wind speed sensor at the corresponding height position of each discrete segment as the local wind speed value, and read the wind direction data output by the wind direction sensor;

[0076] Calculate the difference in local wind speed between adjacent discrete segments. When the difference exceeds a preset gradient threshold, insert a transition segment between the two discrete segments, perform linear interpolation on the wind speed value of the transition segment, and calculate the wind pressure of the transition segment.

[0077] The bending moment contribution component is obtained by projecting the wind pressure of each discrete segment onto the central axis of the tower. The bending moment contribution components of all discrete segments are accumulated along the height direction to obtain the cumulative bending moment distribution curve. The position with the maximum gradient in the cumulative bending moment distribution curve is identified as the stress concentration height.

[0078] The center height coordinates, wind pressure, and bending moment contribution components of each discrete segment are constructed into a three-dimensional data array. The discrete segments corresponding to the stress concentration height are marked and arranged in ascending order of center height coordinates to form a wind pressure distribution map containing stress concentration location information.

[0079] During wind load monitoring, tower structural parameters include total height, cross-sectional dimension variation, material parameters, and connection node locations. Based on the total tower height and sensor spacing, the tower is divided into segments at preset intervals, typically ranging from 5 to 15 meters, with the specific value determined by the tower height and structural complexity. When the tower height exceeds 80 meters, the preset interval is preferably 8 to 12 meters; when the tower height is between 40 and 80 meters, the preset interval is preferably 5 to 8 meters. The segmentation process begins at the tower base and proceeds segment by segment along the height direction. The starting and ending positions of each discrete segment are determined by its vertical coordinates in the global coordinate system.

[0080] The center height coordinates of each discrete segment are obtained by taking the arithmetic mean of the starting and ending coordinates. The calculation of the windward area needs to consider the tower's cross-sectional shape and component arrangement. For a standard square cross-section tower, the windward area equals the product of the discrete segment's height and cross-sectional width, multiplied by a fill factor. The fill factor reflects the proportion of the solid portion of the tower's frame structure to the projected area, typically between 0.25 and 0.45. For steel truss towers, the fill factor is generally between 0.3 and 0.35. For conical towers where the cross-sectional dimensions vary along the height, the cross-sectional widths at the top and bottom of the discrete segment need to be calculated separately, and the average value is taken as the equivalent width of that segment. This average is then multiplied by the discrete segment's height to obtain the windward area.

[0081] Wind speed sensors are installed at corresponding heights in each discrete segment. The wind speed data output by the sensors is digitally filtered to serve as the local wind speed value at that height. The digital filtering uses a moving average window with a window length set between 10 and 30 seconds to eliminate high-frequency fluctuations from instantaneous gusts. Wind direction sensors are typically installed at the top of the tower or the base of the boom. The output wind direction data is represented by a clockwise rotation angle with true north as the 0-degree reference. The wind direction data is used to determine the direction of wind pressure. When the wind direction angle is within ±45 degrees of the normal direction of a certain face of the tower, that face is considered the main windward face.

[0082] The wind pressure in each discrete segment is calculated using Bernoulli's wind pressure formula, which states that wind pressure equals the product of air density, the square of the local wind speed, the windward area, and the drag coefficient, divided by 2. The air density is corrected for on-site temperature and atmospheric pressure, and is taken as 1.225 kg / m³ under standard conditions. 3The density increases by approximately 4% when the temperature drops by 10 degrees Celsius. The drag coefficient is related to the tower's cross-sectional shape and surface roughness; typical drag coefficients for square-section steel trusses range from 1.3 to 1.8, while those for circular sections range from 0.6 to 0.8. Wind pressure calculations are in Newtons. For a standard tower with a height of 80 meters and a cross-sectional width of 2 meters, the wind pressure in a single discrete section is approximately 3000 to 5000 Newtons under a wind speed of 15 meters per second.

[0083] The difference in local wind speed values ​​between adjacent discrete segments reflects the rate of change of wind speed along height. The wind speed gradient is obtained by calculating the difference in local wind speed values ​​between the upper and lower discrete segments and dividing it by the difference in their center height coordinates. A preset gradient threshold is set between 0.15 and 0.30. When the actual gradient exceeds this threshold, it indicates a significant abrupt change in wind speed in the area, possibly caused by building obstruction, terrain influence, or atmospheric turbulence. In this case, a transition segment is inserted between the two discrete segments, and the center height coordinate of the transition segment is taken as the arithmetic mean of the center heights of the two adjacent discrete segments.

[0084] The wind speed value in the transition section is calculated using linear interpolation. Let the local wind speed value of the lower discrete section be... The local wind speed value of the upper discrete segment is Wind speed value in the transition section according to Confirmed. The windward area of ​​the transition section is taken as the average of the windward areas of two adjacent discrete sections, and the drag coefficient remains consistent with that of the adjacent sections. The calculation method for wind pressure in the transition section is the same as that for the conventional discrete section; the interpolated wind speed value is substituted into the wind pressure formula. After inserting the transition section, the wind pressure distribution along the height of the tower is described more precisely, avoiding calculation errors caused by sudden changes in wind speed.

[0085] The wind pressure in each discrete segment acts as a concentrated horizontal force at the center height of that segment. Projecting the wind pressure onto the tower's central axis yields the bending moment contribution component generated by this force at the tower base. This bending moment contribution component equals the wind pressure multiplied by the center height coordinate of that discrete segment, representing the contribution of the wind load in that segment to the bending moment at the tower root. The bending moment contribution components of all discrete segments are accumulated from bottom to top along the height direction to form a cumulative bending moment distribution curve. The vertical axis of this curve represents height, and the horizontal axis represents the cumulative bending moment value generated by all wind pressures from the tower base to that height.

[0086] The gradient of the cumulative bending moment distribution curve reflects the rate of increase of the bending moment along the height. By numerically differentiating the curve and dividing the increment of the cumulative bending moment between adjacent height points by the increment of height, the local gradient value is obtained. The location of the maximum gradient corresponds to the region with the steepest slope of the curve; the bending moment increases most rapidly at this point, indicating that the wind load effect is most significant and the stress is most concentrated at this height. The method to identify the location of the maximum gradient is to traverse all gradient values, find the point with the largest value, and record the corresponding height coordinates as the stress concentration height. Stress concentration heights typically occur at locations of abrupt changes in the tower cross-section, at connection nodes, or in areas with large wind speed gradients.

[0087] When constructing the three-dimensional data array, the first dimension is the discrete segment index, numbered sequentially from bottom to top; the second dimension contains three fields, storing the center height coordinates, wind pressure, and bending moment contribution components, respectively; the third dimension is a labeling information used to indicate whether the discrete segment is located at a stress concentration height. In the specific implementation, for the discrete segment corresponding to the stress concentration height, a special identifier is added to its data record. The identifier can be a Boolean variable, assigned a value of true when the segment is a stress concentration segment, and false otherwise.

[0088] The data array is arranged in ascending order of its center height coordinates to ensure that the wind pressure distribution map is presented sequentially from the tower base to the top. The arrangement process uses a quicksort algorithm, with the center height coordinates as the sorting key. After arrangement, the data array is traversed, and the wind pressure values ​​for each discrete segment are extracted to generate a line graph of wind pressure variation with height. Simultaneously, stress concentration locations are marked, and these discrete segments are highlighted in the graph with different colors or symbols to facilitate rapid identification of hazardous areas by monitoring personnel.

[0089] The wind pressure distribution map not only contains numerical data but also integrates spatial location information and mechanical characteristics. The map allows for a direct observation of the wind load distribution along the tower's height, enabling the assessment of any localized overloading. When the wind pressure in a discrete segment is significantly higher than that in adjacent segments, and this segment is also marked as a stress concentration location, it indicates a higher structural risk in that area. The map data is stored in a standardized format, facilitating subsequent input into deep neural networks for pattern recognition and instability prediction.

[0090] The generation cycle of the entire wind pressure distribution map is synchronized with the sensor sampling frequency, typically updating every 1 to 5 seconds. The real-time updated map reflects the dynamic changes in wind load, capturing the impact of transient events such as gusts and sudden wind direction changes on the tower's stress. Historical map data is stored in time series for analyzing the temporal evolution of wind load, providing rich sample data for training deep neural networks.

[0091] In practical engineering applications, wind pressure distribution maps and strain data are analyzed collaboratively. The strain sensor output values ​​at stress concentration heights should match the high-stress areas predicted by the wind pressure distribution map. Significant discrepancies between the two may indicate sensor malfunction, structural damage, or deviations in the calculation model parameters, necessitating a secondary verification process. This involves adding temporary sensors or conducting on-site testing to verify data accuracy. The wind pressure distribution map, as the core data structure of the monitoring system, permeates the entire AI monitoring process, providing fundamental information support for subsequent load decomposition, instability prediction, and early warning decisions.

[0092] Obtain strain calibration data under no-load and full-load conditions, represent the strain data as a weighted combination of no-load and full-load strain calibration data, solve for the weighting coefficients by minimizing the fitting error, and extract the wind load strain components from the strain data based on the weighting coefficients, including:

[0093] The crane is controlled to be in no-load and full-load states in a windless environment. The output of the strain sensor is collected and marked as no-load strain calibration data and full-load strain calibration data.

[0094] Strain data of different combinations of lifting weight and wind speed under actual working conditions were collected. A three-dimensional response surface of strain data, lifting weight and wind speed was established. The strain value corresponding to zero lifting weight on the three-dimensional response surface was extracted as the pure wind load strain benchmark, and the strain value corresponding to zero wind speed was extracted as the pure lifting weight strain benchmark.

[0095] The correlation coefficient between the pure suspended load strain benchmark and the no-load strain calibration data and the full-load strain calibration data is calculated as the initial weighting coefficient;

[0096] A search space is constructed with the initial weight coefficients as the center. The weight coefficients in the search space are traversed. The unloaded strain calibration data and the full-load strain calibration data are weighted and combined according to each weight coefficient to obtain the reconstructed strain. The fitting error between the reconstructed strain and the pure suspended strain benchmark is calculated. The weight coefficient with the smallest fitting error is selected as the optimal weight coefficient.

[0097] The wind load strain components are obtained by weighting the unloaded strain calibration data and the full-load strain calibration data according to the optimal weighting coefficient, and then subtracting the weighted combination result from the strain data.

[0098] Strain calibration in a windless environment is fundamental for extracting the strain components of wind load. The calibration experiment was conducted in calm weather with wind speeds below 0.5 m / s, after the crane had ceased operation. First, the crane hook was completely unloaded, ensuring the total weight of the hook and lifting equipment did not exceed 2% of the rated load. At this point, the values ​​recorded by the strain sensors primarily reflect the bending strain caused by the boom's own weight. With the boom remaining horizontal and its amplitude fixed, strain data was continuously collected for 30 minutes at a sampling frequency of 10 Hz. Outliers with fluctuations exceeding ±3% of the average value were removed, and the average value was taken as the no-load strain calibration data. Subsequently, a counterweight corresponding to the crane's rated lifting capacity is loaded onto the hook. The mass error of the counterweight must be controlled within ±0.5%. The boom is kept horizontal and its amplitude is consistent with the no-load calibration. Strain data is acquired and processed using the same acquisition duration and frequency to obtain the full-load strain calibration data. Throughout the calibration process, it is essential to ensure that the change in elevation angle displayed by the boom angle sensor does not exceed 0.2° to avoid calibration errors caused by attitude changes.

[0099] Strain data acquisition under actual operating conditions needs to cover multiple combinations of operating conditions within the crane's working range. Five different lifting load levels were selected, starting from no load and increasing in increments of 20%, 40%, 60%, 80%, and 100% of the rated load. Strain data was collected under natural wind conditions of 1 m / s, 3 m / s, 5 m / s, 7 m / s, and 9 m / s for each lifting load level. To obtain stable wind speed conditions, experiments were conducted during periods of relatively small wind speed fluctuations, based on real-time wind speed data released by the meteorological department. Data acquisition time for each operating condition point was no less than 10 minutes. The collected strain data... By lifting weight and wind speed Discrete data points are established, and a three-dimensional response surface is constructed using the bicubic spline interpolation method. Extract the weight from the curved surface. All corresponding strain values, primarily caused by wind load, are arranged according to wind speed to form a pure wind load strain reference sequence. Similarly, wind speed is extracted. The corresponding strain values, which mainly reflect the strain generated by the suspended load, form a pure suspended load strain reference sequence. .

[0100] The initial weighting coefficients were calculated based on the correlation analysis between the pure load-bearing strain benchmark and the calibration data. For each data point in the pure load-bearing strain benchmark sequence... Construct its calibration data with unloaded strain. and full-load strain calibration data The linear combination relationship. Calculate the pure suspension strain reference sequence and... Pearson correlation coefficient and with correlation coefficient The correlation coefficients are normalized to obtain the initial weight coefficients. and Satisfying the normalization condition The initial weighting coefficient reflects the basic proportional relationship between the contributions of no-load and full-load calibration data to actual strain under different lifting conditions.

[0101] The weight coefficients are optimized using a strategy combining grid search and gradient descent. Initial weight coefficients are used... Construct a search space centered on [the search space]. The step size is set to 0.01. Each candidate weight coefficient is traversed within the search space. Calculate accordingly The reconstructed strain is obtained by weighting and combining the no-load strain calibration data and the full-load strain calibration data. The reconstructed strain is compared with the pure load-bearing strain reference sequence, and the root mean square error is calculated. ,in This represents the number of data points in the pure suspension strain benchmark sequence. After iterating through all candidate weighting coefficients, the one that makes the weighting coefficients most suitable is selected. Weight coefficient combination that reaches the minimum value The optimal weighting coefficients were used as the basis for cross-validation. To verify the stability of the optimization results, the optimal weighting coefficients were substituted into the strain data under different amplitude conditions to ensure that the fitting error remained within 5% under each condition.

[0102] The extraction of wind load strain components is based on the inverse decomposition of the strain superposition principle. This involves extracting strain data collected under actual operating conditions. Considered as the strain component of the suspended weight Wind load strain components The linear superposition, i.e. By using optimal weighting coefficients to weight and combine the no-load strain calibration data and the full-load strain calibration data, the estimated values ​​of the suspended strain components are obtained. This estimate reflects the strain response caused solely by the load under the current lifting conditions. Subtracting this estimate from the actual measured strain yields the wind load strain components. To eliminate high-frequency noise caused by instantaneous gusts, the extracted wind load strain components were low-pass filtered with a cutoff frequency of 0.5 Hz to retain low-frequency components that reflect the trend of wind load changes.

[0103] In practical applications, corresponding calibration databases need to be established for different boom amplitudes and boom lengths. Since boom extension or luffing alters the structural stiffness distribution, calibration data for a single amplitude is not applicable to all operating conditions. Therefore, five typical boom amplitudes commonly used in cranes are selected: minimum amplitude, 25% of maximum amplitude, 50% of maximum amplitude, 75% of maximum amplitude, and maximum amplitude. The no-load and full-load calibration process is repeated for each amplitude. For intermediate amplitudes encountered in actual operation, the corresponding calibration data is obtained by linear interpolation of the calibration data of adjacent typical amplitudes. and Meanwhile, considering the temperature drift characteristics of strain sensors, the ambient temperature is recorded in the calibration data. When the difference between the actual operating temperature and the calibration temperature exceeds 10°C, the calibration data is corrected according to the temperature compensation coefficient of the strain sensor. The correction formula is as follows: ,in The temperature coefficient of the strain sensor, This refers to the temperature difference.

[0104] To improve the real-time performance of wind load strain component extraction, a sliding window mechanism is employed for online calculation. A 60-second time window is set, with strain data updated every second. The data within the window is weighted and averaged before wind load strain component extraction. The weighting coefficients within the window use an exponential decay method, with data closer to the current moment having a higher weight, ensuring the extraction results quickly reflect actual wind load changes. Simultaneously, an anomaly detection mechanism for wind load strain components is established. When a sudden change occurs in the extracted wind load strain component with a rate of change exceeding three times the normal fluctuation range, a data verification process is triggered to re-check the sensor status and the validity of the weighting coefficients, avoiding misjudgments caused by sensor malfunctions or invalid calibration data.

[0105] Strain data were collected under actual operating conditions with different combinations of lifting weight and wind speed. A three-dimensional response surface was established based on the strain data and the relationship between lifting weight and wind speed, including:

[0106] Set the range of load variation and wind speed variation. Within the range of load variation, select multiple load sampling points evenly, and within the range of wind speed variation, select multiple wind speed sampling points evenly. Combine the load sampling points and wind speed sampling points by Cartesian product to obtain the set of sampling working condition points.

[0107] The crane is controlled to execute the lifting weight and wind speed conditions corresponding to each sampling point in the sampling point set one by one, and the output values ​​of the strain sensor corresponding to each sampling point are collected to establish a mapping relationship between the sampling points and the output values ​​of the strain sensor.

[0108] The sampling points in the sampling point set that have missing strain sensor output values ​​are marked as missing points. The neighboring points around the missing points that have collected strain sensor output values ​​are extracted. The interpolated strain value of the missing point is calculated by weighted interpolation based on the strain sensor output values ​​of the neighboring points and the distance between the neighboring points and the missing point.

[0109] The load value of the sampling point is used as the horizontal axis coordinate, the wind speed value of the sampling point is used as the vertical axis coordinate, and the output value of the strain sensor and the interpolated strain value corresponding to the sampling point are used as the vertical axis coordinate to construct a three-dimensional coordinate point cloud.

[0110] A surface fitting operation is performed on a 3D coordinate point cloud to generate a 3D response surface.

[0111] In actual operating environments, the strain experienced by a crane is influenced by two main factors: the load being lifted and the wind speed, with a complex coupling relationship between them. To accurately describe this multi-factor coupling effect, a three-dimensional response model of the load, wind speed, and strain needs to be established.

[0112] Before starting data collection, the reasonable range of variation for the load and wind speed must first be determined. The load variation range is typically set to 10% to 100% of the crane's rated lifting capacity, ensuring coverage of the entire working range from light to full load. The wind speed variation range is determined based on the meteorological conditions permitted for crane operation, generally set from 0 to the maximum wind speed allowed for crane operation. For large tower cranes, this upper limit is typically 12 to 15 m / s. Within the load variation range, 15 to 25 load sampling points are selected at equal intervals to ensure even distribution and sufficient density. For example, for a crane with a rated lifting capacity of 60 tons, a load sampling point can be set every 3 tons within the range of 6 to 60 tons, resulting in a total of 19 load sampling points. Within the wind speed variation range, 12 to 20 wind speed sampling points are also selected at equal intervals, typically set at 0.5 to 1 m / s intervals.

[0113] Perform a Cartesian product operation on all load sampling points and wind speed sampling points to generate a complete two-dimensional load condition sampling matrix. Each element in this matrix represents an independent sampling load point, containing a specific combination of load and wind speed values. If the number of load sampling points is... The number of wind speed sampling points is The total number of sampling points is This combination method ensures complete coverage of the sampling space, providing ample data support for subsequent surface fitting.

[0114] During actual sampling, strict control of test conditions is necessary to ensure data quality. For each sampling point, the load on the crane hook is first adjusted to the target lifting weight, and the load accuracy is verified using a load cell; the error should be controlled within ±2% of the target lifting weight. Then, wait for the wind speed conditions to meet the requirements. When the real-time monitored wind speed value stabilizes within ±0.3 m / s of the target wind speed value for more than 30 seconds, the wind speed conditions are considered to meet the sampling requirements. After meeting these two conditions, the strain sensor output value is continuously recorded for a sampling duration of no less than 60 seconds, with a sampling frequency set to 50Hz to 100Hz. The arithmetic mean of the strain data within this period is taken as the strain sensor output value corresponding to that sampling point, and the standard deviation of the data is recorded to evaluate measurement stability.

[0115] During actual sampling, due to the uncontrollability of meteorological conditions or equipment operation limitations, some sampling points may fail to complete data acquisition. Sampling points that fail to obtain valid strain sensor output values ​​are marked as missing points. For each missing point, a search is conducted within the sampling point set for neighboring points that have successfully acquired data. The selection of neighboring points is based on the Euclidean distance criterion. The distance between the missing point and other acquired points in the two-dimensional space of load-wind speed is calculated, and the 6 to 8 closest points are selected as neighboring points. The distance calculation requires normalization of the load and wind speed to avoid deviations caused by dimensional differences. The normalization formula is to divide the load value by the maximum load value and the wind speed value by the maximum wind speed value, ensuring that both values ​​fall within the range of 0 to 1.

[0116] The strain values ​​for missing working conditions are calculated using an inverse distance weighted interpolation method. The core idea of ​​this method is that the closer a neighboring working condition is to the missing working condition, the greater its contribution weight to the interpolation result. Specifically, the... The weight of each neighboring operating point The distance to the missing working point The distances are inversely proportional. To avoid numerical singularities when the distance is zero, an inverse distance weighting method with a power of 2 is used. The sum of the weights of all neighboring working points is normalized to 1 to ensure the reasonableness of the interpolation results. The interpolated strain value for a missing working point is equal to the sum of the products of the strain values ​​of each neighboring working point and their corresponding weights. This interpolation method can better maintain the continuity and smoothness of the strain field and avoid abrupt changes at the locations of missing data.

[0117] After acquiring or interpolating all sampled load points, a three-dimensional coordinate system is established for data visualization and surface construction. The horizontal axis is defined as the load axis, the vertical axis as the wind speed axis, and the interpolated axis as the strain axis. For each load point in the sampled load point set, its load value, wind speed value, and corresponding strain sensor output value or interpolated strain value are used as the horizontal, vertical, and interpolated coordinates of that point in three-dimensional space, respectively. All sampled load points are mapped to this three-dimensional coordinate system, forming a discrete three-dimensional coordinate point cloud. This point cloud realistically reflects the spatial distribution characteristics of strain variation with load and wind speed.

[0118] A surface fitting operation is performed on the 3D coordinate point cloud to construct a continuous 3D response surface. The surface fitting employs either polynomial fitting or spline interpolation methods. The polynomial fitting method expresses the strain values ​​as a polynomial function of the load and wind speed; the polynomial order is chosen based on the data complexity, typically a quadratic or cubic polynomial. The fitting process uses the least squares method to determine the coefficients of each polynomial term, minimizing the sum of squared residuals between the fitted surface and the actual measurement points. The spline interpolation method, on the other hand, constructs locally smooth piecewise functions while ensuring the surface passes through all measurement points; this method is more adaptable to local variations in the data. Regardless of the fitting method used, the fitting results need to be verified by calculating the root mean square error between the fitted surface and the actual measurement points. This error should be less than 5% of the strain measurement value to ensure that the accuracy of the fitted surface meets the requirements of engineering applications.

[0119] The generated 3D response surface can intuitively display the overall trend and local characteristics of strain variation with load and wind speed. When the load is constant, observing the surface morphology along the wind speed direction reveals that the strain value increases monotonically with increasing wind speed, with the growth rate typically faster in high-wind-speed regions, reflecting the nonlinear influence of wind load on structural strain. When the wind speed is constant, observing the surface morphology along the load direction also shows an increasing strain value with increasing load, but the slope of the growth curve is relatively stable, reflecting the linear characteristics of the gravity load effect. Changes in surface curvature reveal the coupling effect between load and wind speed. Under certain specific combinations of operating conditions, the surface may exhibit local bulges or depressions, indicating an abnormal strain rate of change near that point, requiring close attention in actual operations.

[0120] The established three-dimensional response surface is not only used for strain prediction but also serves as a reference for subsequent extraction of wind load strain components. For any measured strain value, the corresponding load-wind speed contour lines can be found on the three-dimensional response surface, thus allowing for the inference of the contribution of wind load under the current operating conditions. This response surface can also be used to optimize crane operation plans. By searching the surface for feasible operating conditions that satisfy strain safety constraints, it provides operators with maximum safe lifting weight recommendations under different wind speed conditions, effectively avoiding the risk of structural instability due to overloading or excessive wind load.

[0121] The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network. Based on historical instability accident samples, the similarity between the current state and the instability mode is output, including:

[0122] The wind pressure distribution map and the wind load strain components are time-aligned to construct a nonlinear coupled response function containing time lag terms, thus obtaining the wind pressure-strain coupled response field. The coupled evolution trajectory sequence is obtained by sliding decomposition of the wind pressure-strain coupled response field along the time dimension.

[0123] The coupled evolution trajectory sequence is input into the spatial coding subnetwork and temporal coding subnetwork of the deep neural network to extract spatial gradient features and evolution trend features, respectively. The spatial gradient features and evolution trend features are then input into the physical constraint module. The physical constraint module applies consistency constraints based on the mechanical equilibrium relationship between wind pressure and strain to obtain constraint feature vectors.

[0124] A coupled consistency loss function containing wind pressure-driven error terms and strain inversion error terms is constructed to train a deep neural network, so that the constraint feature vector satisfies the minimum bidirectional prediction error of wind pressure and strain.

[0125] The current coupled evolution trajectory sequence is dynamically time-warped and matched with the coupled evolution trajectory sequence of historical instability accident samples to calculate the trajectory matching distance. The instability mode with the smallest trajectory matching distance is selected as the target instability mode.

[0126] Calculate the path deviation and evolution trend consistency coefficient between the current coupled evolution trajectory sequence and the target instability mode trajectory sequence, and calculate the similarity based on the path deviation and evolution trend consistency coefficient.

[0127] Before performing instability pattern recognition using deep neural networks, refined time alignment processing of the wind pressure distribution map and wind load strain components is required. Due to the physical time lag in the transmission of wind pressure to structural strain, a time lag correction model needs to be constructed. Although the collected wind speed and strain data are recorded synchronously at a frequency of 10Hz, there is a transmission delay of 3 to 8 seconds between the wind load acting on the tower and the boom strain response. The time lag parameters of the wind pressure and strain sequences are calculated using cross-correlation analysis to determine the optimal alignment time window. Specifically, the wind pressure sequence is slid in 0.1-second increments, and the Pearson correlation coefficient between each slid position and the strain sequence is calculated. The slid value corresponding to the maximum correlation coefficient is selected as the time lag correction value. After time alignment, a nonlinear coupled response function is constructed to describe the dynamic relationship between wind pressure and strain. This function uses a kernel function expansion form and includes an immediate response term and a cumulative lag term, with the cumulative lag term covering the historical wind pressure influence of the previous 15 seconds. Wind pressure data from different times were weighted and fused using a Gaussian radial basis function kernel, with a kernel bandwidth parameter set to 2.5 seconds to ensure both instantaneous impacts and sustained effects were captured. The resulting wind pressure-strain coupled response field comprises three elements: spatial dimension, temporal dimension, and response intensity, forming a three-dimensional data tensor structure. A sliding decomposition was performed along the time dimension with a window length of 30 seconds and a sliding step size of 5 seconds. Data within each time window constitutes a coupled evolution trajectory segment, and consecutive segments form a complete coupled evolution trajectory sequence.

[0128] The coupled evolutionary trajectory sequence is input into a dual-channel encoding architecture of a deep neural network. The spatial encoding subnetwork adopts a convolutional neural network structure, containing four convolutional layers with kernel sizes of 7×7, 5×5, 3×3, and 3×3, and channel numbers of 32, 64, 128, and 256, respectively. The first convolutional layer extracts the distribution gradient of wind pressure at different heights on the tower, the second layer captures the pressure difference patterns between discrete segments, and subsequent convolutional layers gradually integrate global spatial features. Each convolutional layer is followed by a batch normalization layer and a modified linear unit activation function, and the pooling layer uses 2×2 max pooling to reduce spatial resolution. The temporal encoding subnetwork adopts a long short-term memory network structure, containing three recurrent units with hidden state dimensions of 128, 256, and 512, respectively. This network processes the evolutionary sequence within a 30-second time window, capturing the rate of change, acceleration, and periodic fluctuation characteristics of wind load strain. The long short-term memory units, through forget gates, input gates, and output gates, can both retain long-term cumulative effects and respond to short-term abrupt signals. The spatial gradient feature output by the spatial coding subnetwork is a 256-dimensional vector, describing the spatial heterogeneity of wind pressure distribution; the evolution trend feature output by the temporal coding subnetwork is a 512-dimensional vector, characterizing the dynamic evolution of strain response.

[0129] The physical constraint module receives spatial gradient features and evolution trend features, and applies consistency constraints based on the mechanical equilibrium relationship between wind pressure and strain. This module first maps the two types of features to a unified 512-dimensional physical space through fully connected layers. In this physical space, wind pressure features should be able to predict strain features, and vice versa. A bidirectional prediction network is constructed: the forward prediction network uses wind pressure features as input to predict strain features, and the backward prediction network uses strain features as input to predict wind pressure features. Both prediction networks contain three fully connected layers, with intermediate layers of dimensions 256 and 128, and the output layer dimension matching the target feature dimension. By comparing the deviation between the predicted features and the actual features, the physical consistency residual is calculated. The original features and the consistency residual are fused through a residual connection structure to obtain a constraint feature vector corrected by physical constraints. This vector has a dimension of 512, simultaneously satisfying spatial distribution rationality and temporal evolution continuity.

[0130] The deep neural network is trained using a coupling consistency loss function, which includes a wind pressure-driven error term and a strain inversion error term. The wind pressure-driven error term measures the accuracy of predicting strain features from wind pressure features and is defined as the mean square error between the predicted strain features and the actual strain features. The strain inversion error term measures the accuracy of inverting wind pressure features from strain features and is defined as the mean square error between the inverted wind pressure features and the actual wind pressure features. The total loss function is a weighted sum of the two terms, with the wind pressure-driven error weight set to 0.6 and the strain inversion error weight set to 0.4. This weighting is determined based on the physical mechanism where wind pressure is the primary driving factor and strain is the response. The training process uses the Adam optimizer with an initial learning rate of 0.001, which decays to 0.8 times every 20 iterations. The training dataset contains 300 sets of data under normal operating conditions and 120 sets under near-instability critical conditions. The training runs for 150 epochs, with a random batch size of 16 samples selected for gradient updates in each epoch. This training strategy ensures that the constrained feature vectors can accurately reflect wind pressure input and precisely match strain output in physical space, thereby minimizing the bidirectional prediction error of wind pressure and strain.

[0131] In the instability mode identification stage, the coupled evolution trajectory sequence of the current operating state is dynamically time-warped and matched with the trajectory sequences in the historical instability accident sample library. The historical instability accident sample library contains 35 instability cases of different types, covering various instability modes such as tower overturning, boom breakage, and foundation settlement. Each mode contains a complete coupled evolution trajectory from 60 to 120 seconds before instability. The dynamic time warping algorithm can handle the problems of inconsistent time series lengths and local time distortion by constructing a cumulative distance matrix to find the optimal matching path. The elements of the cumulative distance matrix represent the local distance between the current trajectory at time i and the historical trajectory at time j. The local distance uses Euclidean distance to measure the difference in the coupled response field. The dynamic programming algorithm recursively moves from the lower left corner to the upper right corner of the matrix. At each step, it can choose to move upward, to the right, or to the upper right, and the choice of movement direction minimizes the cumulative distance. The sum of the cumulative distances corresponding to the optimal matching path is the trajectory matching distance. The trajectory matching distance is calculated for each of the 35 historical instability samples, and the instability mode corresponding to the sample with the smallest distance value is selected as the target instability mode. If the sample with the smallest matching distance is a tower lateral overturning accident, then the current state evolution trend is determined to be closest to this instability mode.

[0132] After determining the target instability mode, a refined similarity index between the current trajectory and the target mode trajectory is further calculated. The path deviation value quantifies the geometric deviation between the two trajectories in the coupled response space, defined as the square root of the arithmetic mean of the sum of squared local distances between all corresponding point pairs on the dynamic time-warped matching path. This index reflects the difference in the overall shape of the trajectories; a smaller deviation value indicates a more similar trajectory shape. The evolution trend consistency coefficient assesses the synchronization of the change direction and rate of the two trajectories, obtained by calculating the cosine similarity of the evolution velocity vectors at corresponding moments. The evolution velocity vector is calculated by the difference of the coupled response field at adjacent moments, containing two components: the rate of change of wind pressure and the rate of change of strain. The evolution trend consistency coefficient is obtained by averaging the cosine similarity of the velocity vectors of all corresponding point pairs on the matching path. This coefficient ranges from -1 to 1, where 1 indicates completely unidirectional evolution and -1 indicates completely inverse evolution. The similarity score comprehensively considers both path deviation and evolutionary trend consistency coefficient. The calculation formula is a weighted geometric mean of the evolutionary trend consistency coefficient and the inverse of the path deviation, with the specific weights determined through historical data statistical optimization to be 0.55 and 0.45. This similarity index focuses on both the static morphological similarity of the trajectory and the consistency of dynamic evolutionary direction, accurately judging the likelihood and urgency of the current state developing towards the target instability mode. A higher similarity score indicates a closer resemblance between the current operating state and historical instability modes, and a greater risk of structural instability, providing a quantitative basis for subsequent critical moment prediction and safety margin calculation.

[0133] The evolution curve is obtained by nonlinearly fitting the numerical changes of similarity over time. Based on the evolution curve, the critical moments when the similarity reaches the instability threshold are determined, including:

[0134] The similarity scores are collected continuously at multiple times according to a preset sampling period and arranged in chronological order to form a similarity time series;

[0135] The similarity time series is segmented by a sliding window and the rate of change of similarity within each segment is calculated. The time and similarity corresponding to the segments with positive rate of change are extracted to form the unstable trend data.

[0136] The similarity between adjacent sample points in the unstable trend data is subjected to second-order difference operation to obtain a second-order difference value sequence. The absolute value of each second-order difference value in the second-order difference value sequence is normalized and used as the fitting weight coefficient of the corresponding sample point.

[0137] A polynomial function is used to perform weighted curve fitting on the sample points in the unstable trend data to obtain the initial evolution curve and calculate the fitting residual of each sample point. When the fitting residual exceeds the preset residual threshold, the degree of the polynomial function is increased by a preset step size value and the sample points are re-fitted with weighted curve to obtain the optimized evolution curve.

[0138] The instability threshold is substituted into the time variable of the optimization evolution curve solution as the critical time when the similarity reaches the instability threshold.

[0139] In actual crane operation monitoring scenarios, the acquired similarity data is not a static value, but a dynamic process quantity that changes over time. To accurately capture the instability evolution process and precisely predict critical moments, in-depth analysis of the temporal variation characteristics of similarity is required. Specifically, a preset sampling period is set between 2 and 10 seconds to ensure that data redundancy is avoided while capturing dynamic changes. Within this sampling period, similarity values ​​are continuously recorded for the current moment and several previous moments, arranged chronologically to form a similarity time series. The length of this time series depends on the monitoring time window, typically maintaining the most recent 30 to 100 sampling points. This ensures sufficient sample size for statistical analysis while avoiding interference from premature historical data in judging the current state.

[0140] To identify data segments with genuine instability trends from similarity time series, a sliding window technique is introduced for local feature extraction. The length of the sliding window is set to 5 to 15 sampling points, and the step size is set to 1 to 3 sampling points. For the similarity data within each window, the numerical difference between its start and end points is calculated, and then divided by the time span contained within the window to obtain the average rate of change corresponding to that window. After traversing the entire similarity time series, a series of rate of change values ​​corresponding to each window are obtained. Among these rates of change, segments with positive values ​​are selected, indicating that the similarity shows an upward trend, meaning that the crane's state is approaching an instability mode. The time and similarity values ​​corresponding to these upward trend segments are extracted, and stationary segments and declining segments with negative or near-zero rates of change are removed to form an instability trend dataset. Each sample point in this dataset contains information in two dimensions: time coordinates and similarity coordinates.

[0141] Based on the instability trend data, the accelerating characteristics of similarity changes are further analyzed. For adjacent sample points arranged in chronological order in the instability trend data, the first-order difference of their similarity is calculated, which is the similarity of the later sample point minus the similarity of the earlier sample point. Based on the first-order difference sequence, the difference operation is performed again on adjacent elements to obtain the second-order difference value sequence. The second-order difference value reflects the change in the rate of similarity change; the larger the absolute value, the more drastic the change trend at that position, and the more important it is in characterizing the evolution law. To convert the second-order difference values ​​into fitting weights, the absolute value of all elements in the second-order difference value sequence is taken, and then the maximum absolute value in the sequence is found. The absolute value of the second-order difference value corresponding to each sample point is divided by this maximum absolute value to obtain the normalized weight coefficient between 0 and 1. Sample points with drastic changes receive larger weight coefficients, while sample points with gradual changes receive smaller weight coefficients. This weighting strategy makes the subsequent curve fitting focus more on the key turning points in the instability evolution process.

[0142] A polynomial function is used as the mathematical model for the evolution curve, and the similarity is set as... , at all times The polynomial function is expressed as ,in Let the degree be a polynomial. The coefficients are to be determined. In the initial stage, the polynomial degree is set to 2 or 3, and a weighted least squares fit is performed on the sample points in the unstable trend data. Specifically, a weighted residual sum of squares function is constructed. ,in This represents the total number of sample points in the unstable trend data. For the first Fitting weight coefficients for each sample point and These represent the similarity and time of the sample point, respectively. Solving for... The minimum coefficient combination yields the initial evolution curve. Substituting the time of each sample point in the instability trend data into the initial evolution curve, the predicted similarity value is calculated. The difference between this predicted value and the actual similarity value is the fitting residual.

[0143] The absolute value of the fitting residuals for all sample points is taken, and the maximum value is calculated. This maximum fitting residual is then compared with a preset residual threshold. The preset residual threshold is usually set based on the numerical range of similarity, for example, 0.05 to 0.1. If the maximum fitting residual does not exceed the preset residual threshold, the initial evolution curve is considered sufficiently accurate and is used as the optimized evolution curve. If the maximum fitting residual exceeds the preset residual threshold, it indicates that the current polynomial degree cannot adequately describe the complex nonlinear characteristics of the unstable evolution, and the model complexity needs to be increased. In this case, the polynomial degree is increased by a preset step size, which is usually set to 1, i.e., from degree 3 to degree 4, or from degree 4 to degree 5. The weighted least squares fitting process is re-executed using the new polynomial degree, the new fitting residual is calculated, and compared with the preset residual threshold again. This iterative optimization process is repeated until the fitting residual meets the accuracy requirements or the polynomial degree reaches the upper limit. The upper limit is generally set to degree 6 to 8 to avoid overfitting caused by excessively high degrees. The curve that finally meets the accuracy requirements is the optimized evolution curve.

[0144] When predicting the critical moment of instability based on the optimized evolution curve, an instability threshold needs to be pre-set. This instability threshold is determined by analyzing the statistical distribution of similarity in historical instability accident samples, typically taking the lower quartile of the similarity at the time of the historical instability event to ensure sufficient lead time for the warning. Substituting the instability threshold into the expression of the optimized evolution curve, i.e., letting... ,in As the instability threshold, we obtain the time-varying variables. polynomial equations The equation may have multiple real solutions, and we need to select the solutions with reasonable physical meaning. First, we eliminate solutions with values ​​less than the current time, as these represent past moments and have no predictive significance. Among the remaining solutions greater than the current time, we select the one with the smallest value as the critical moment. This moment represents the time when the similarity first reaches the instability threshold under the condition that the current evolutionary trend continues.

[0145] In practical engineering applications, if the evolution curve fails to reach the instability threshold within the foreseeable future, it indicates that the current crane condition is still within a sufficient safety margin, and no critical moment prediction result is output in this case. If the evolution curve shows a rapid upward trend and is about to reach the instability threshold in a short time, the critical moment calculation is quickly completed and transmitted to the subsequent early warning decision-making stage. To improve the robustness of the prediction, a multi-model fusion strategy can be introduced, and other nonlinear models such as exponential functions and logarithmic functions can be used for fitting. The critical moments predicted by each model are weighted and averaged or voted to further reduce the prediction bias of a single model. The entire process of evolution curve fitting and critical moment solution realizes the transformation from discrete similarity data to continuous-time prediction, providing an accurate time benchmark for subsequent safety margin calculation.

[0146] The safety margin is calculated based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, an early warning signal and action restriction command are generated, including:

[0147] Calculate the time difference between the critical moment and the current moment;

[0148] Obtain the current lifting mass, amplitude value, and slewing angular velocity of the crane. Input the lifting mass, amplitude value, and slewing angular velocity into the historical instability sample library for matching. Extract the actual braking duration of the target historical sample record with the highest matching degree. Subtract the actual braking duration from the time difference to obtain the controllable time window.

[0149] Obtain the current wind pressure distribution map, input the current wind pressure distribution map into the historical instability sample library for map shape comparison, and count the proportion of samples with wind load abrupt changes in the candidate historical samples with matching shapes as the abrupt change probability. Based on the abrupt change probability, compress the controllable time window to obtain the safety margin.

[0150] When the safety margin is less than the preset safety value, the safety margin is entered into the historical instability sample library for matching, and the slewing speed limit, lifting speed limit and amplitude speed limit recorded in the reference historical sample with the highest matching degree are extracted.

[0151] The safety margin and mutation probability are encapsulated to generate an early warning signal. The slewing speed limit, hoisting speed limit, and amplitude speed limit are encapsulated to generate a motion restriction command. The early warning signal is sent to the monitoring terminal, and the motion restriction command is sent to the crane controller.

[0152] After predicting the risk of crane instability, the predicted critical moment needs to be translated into actionable control decisions. The time difference between the critical moment and the current moment is calculated; this time difference reflects the theoretical time span from the current state to the potential instability. The time difference is calculated by subtracting the current moment's timestamp from the critical moment's timestamp, yielding a value in seconds. This time difference includes the total time required for the current operating condition to develop until the instability threshold is reached; however, this theoretical value does not consider the response characteristics of the crane actuators or complex on-site factors.

[0153] The system acquires the crane's current lifting mass, amplitude, and slewing angular velocity. The lifting mass is measured in real-time by a load sensor on the hoisting mechanism, which directly reads the wire rope tension. The actual lifting mass is then calculated using pulley system ratio conversion. The amplitude is measured by an angle encoder installed at the boom root, measuring the boom elevation angle. Combined with the boom length, the horizontal distance from the hook center to the slewing center is calculated geometrically. The slewing angular velocity is measured by a rotary encoder on the slewing bearing, recording the angular change per unit time. These three key operating parameters are input into a historical instability sample database for matching. This database stores all past instability events or near-instability cases. Each record includes the lifting mass, amplitude, slewing angular velocity at the time of the event, and the actual braking time from the issuance of the braking command to the crane's complete stop. The matching process uses a weighted Euclidean distance algorithm, comparing the current operating parameters with those of the historical samples one by one. A smaller distance value indicates a higher degree of matching. When calculating the distance, different weighting coefficients are assigned to the lifting mass, amplitude value, and slewing angular velocity. Typically, the weight for lifting mass is set to 0.5, the weight for amplitude value to 0.3, and the weight for slewing angular velocity to 0.2, reflecting the different degrees of influence of each parameter on the risk of instability. The target historical sample with the highest matching degree, i.e., the sample record with the smallest distance value, is extracted, and the actual braking duration of this sample record is read. The actual braking duration refers to the entire process from when the controller receives the braking command to when all moving mechanisms of the crane completely stop and lock. This duration is affected by multiple factors such as mechanical inertia, hydraulic system response delay, and brake action time, and typically varies between 3 and 15 seconds. The controllable time window is obtained by subtracting the actual braking duration from the time difference. The controllable time window represents the effective decision-making time between issuing the control command and when the crane must complete the response action under the current operating conditions. During this period, the operator or automatic control system can adjust the operating parameters to avoid instability.

[0154] The current wind pressure distribution map is obtained, which records the wind pressure values ​​and spatial distribution patterns experienced by discrete sections of the tower at different heights. This map is then input into a historical instability sample database for pattern comparison. Each record in the database contains operating parameters and a sequence of wind pressure distribution maps for a period prior to the instability event. The pattern comparison employs a dynamic time warping algorithm to align the current map with historical maps spatially and calculate similarity. Similarity evaluation metrics include the overall trend of the map curves, peak positions, gradient change rates, and other characteristic parameters. A comprehensive score is used to select candidate historical samples with matching patterns. The threshold for pattern matching is set at a similarity score greater than 0.75; historical samples meeting this condition are included in the candidate set. The percentage of candidate historical samples experiencing sudden wind load changes is calculated as the probability of such changes. A sudden wind load change is defined as a wind pressure increase exceeding 30% or a wind direction change angle greater than 45 degrees within 5 seconds. The candidate historical sample set is traversed, and each sample record is checked to see if a wind condition change meeting the definition of abrupt change occurred within the corresponding time period. The cumulative number of samples with abrupt changes is divided by the total number of candidate samples to obtain the abrupt change probability. The abrupt change probability reflects the likelihood of a drastic change in wind load in the near future under the current wind pressure distribution pattern. The controllable time window is compressed based on the abrupt change probability to obtain a safety margin. The compression factor is set to [value missing]. ,in This represents the probability of mutation. When the mutation probability is 0, the compression factor is 1, and the controllable time window is not compressed; when the mutation probability is 1, the compression factor is 0.2, and the controllable time window is significantly compressed to 20% of its original value. The safety margin is calculated by multiplying the controllable time window by the compression factor, and this value represents the actual available decision time after considering the uncertainty of wind load.

[0155] When the safety margin is less than the preset safety value, the current state is deemed to have an instability risk, requiring the activation of protective measures. The preset safety value is set according to the crane model and operating environment, generally ranging from 8 to 20 seconds, with 12 seconds typically set for large tower cranes. The safety margin is entered into a historical instability sample database for matching. The matching logic searches the database for historical records with the closest safety margin value to the current value. Each historical sample records the speed limiting measures taken under specific safety margin conditions and their actual effects. The slewing speed limit, hoisting speed limit, and amplitude speed limit values ​​from the reference historical sample record with the highest matching degree are extracted. The slewing speed limit specifies the maximum permissible angular velocity of the slewing mechanism, usually in degrees per second. The speed limit decreases as the safety margin decreases; when the safety margin is less than 5 seconds, the slewing speed limit may drop to 10% of the normal speed or even require a complete stop of slewing. The hoisting speed limit specifies the maximum permissible linear velocity of the hoisting mechanism, in meters per minute. When the safety margin is small, the descent speed is prioritized to prevent exacerbation of load sway. The amplitude speed limit specifies the maximum permissible operating speed of the luffing mechanism, also in meters per minute. When the safety margin is insufficient, it is usually required to stop increasing the amplitude and only allow the boom to be slowly retracted to reduce the overturning moment.

[0156] Safety margin and mutation probability are encapsulated to generate early warning signals. These signals use a structured data format, including signal type, timestamp, safety margin value, mutation probability value, and risk level fields. Risk levels are divided into three levels based on the safety margin: low risk (greater than 15 seconds), medium risk (8-15 seconds), and high risk (less than 8 seconds). The early warning signals also include auxiliary information such as instability mode similarity and current operating parameters, facilitating a comprehensive understanding of the site conditions by monitoring personnel. Slewing speed limits, hoisting speed limits, and amplitude speed limits are encapsulated to generate motion restriction commands. These commands use a control protocol format, including command type identifier, target actuator number, speed limit value, and execution priority. A mandatory execution flag is set in the command to ensure the controller responds immediately upon receiving the command, without manual intervention. The early warning signal is sent to the monitoring terminal, which receives the signal via a wireless communication network and displays color and sound alerts to the operator on the human-machine interface, showing a safety margin countdown and suggested operational measures. The system sends motion restriction commands to the crane controller. Upon receiving the command, the controller immediately executes speed limiting control, adjusting the inverter output frequency or hydraulic valve opening to limit the speed of each moving mechanism. Simultaneously, the controller activates real-time monitoring mode to continuously track the actual operating speed of each mechanism, ensuring it does not exceed the speed limit, and feeds back the execution status to the monitoring system to form a closed-loop control system.

[0157] A second aspect of the present invention provides an AI monitoring system for cranes that takes wind load into account, comprising:

[0158] The data acquisition unit is used to collect wind speed and wind direction data through wind speed sensors and wind direction sensors arranged at different height positions on the crane tower, and to collect strain data through strain sensors arranged on the boom.

[0159] The wind pressure calculation unit is used to divide the tower into multiple discrete segments according to its height, calculate the local wind speed value based on the wind speed data of each discrete segment, and calculate the wind pressure of each discrete segment by combining the windward area and wind resistance coefficient, and arrange them according to spatial location to form a wind pressure distribution map.

[0160] The strain analysis unit is used to acquire strain calibration data under no-load and full-load conditions. It represents the strain data as a weighted combination of no-load strain calibration data and full-load strain calibration data. The weight coefficients are solved by minimizing the fitting error, and the wind load strain components are extracted from the strain data based on the weight coefficients.

[0161] The instability prediction unit is used to input the wind pressure distribution map and wind load strain components into a pre-trained deep neural network, and output the similarity between the current state and the instability mode based on historical instability accident samples.

[0162] The critical judgment unit is used to perform nonlinear fitting on the numerical change of similarity over time to obtain the evolution curve, and to determine the critical moment when the similarity reaches the instability threshold based on the evolution curve.

[0163] The early warning generation unit is used to calculate the safety margin based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, it generates an early warning signal and an action restriction command.

[0164] A third aspect of the present invention provides an electronic device, comprising:

[0165] processor;

[0166] Memory used to store processor-executable instructions;

[0167] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0168] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0169] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0170] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A crane AI monitoring method considering wind load, characterized in that, include: Wind speed and wind direction data are collected by wind speed sensors and wind direction sensors placed at different heights on the crane tower, and strain data are collected by strain sensors placed on the boom. The tower is divided into multiple discrete segments according to its height. The local wind speed value is calculated based on the wind speed data of each discrete segment. The wind pressure of each discrete segment is calculated by combining the windward area and wind resistance coefficient. The segments are arranged according to their spatial location to form a wind pressure distribution map. Obtain strain calibration data under no-load and full-load conditions, represent the strain data as a weighted combination of no-load strain calibration data and full-load strain calibration data, solve the weight coefficient by minimizing the fitting error, and extract the wind load strain component from the strain data based on the weight coefficient. The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network, and the similarity between the current state and the instability mode is output based on historical instability accident samples. The evolution curve is obtained by nonlinear fitting of the numerical change of similarity over time. The critical moment when the similarity reaches the instability threshold is determined based on the evolution curve. The safety margin is calculated based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, an early warning signal and action restriction command are generated.

2. The method according to claim 1, characterized in that, The tower is divided into multiple discrete segments according to its height. Local wind speed values ​​are calculated based on wind speed data for each discrete segment. Wind pressure in each discrete segment is calculated by combining the windward area and drag coefficient. These segments are then arranged spatially to form a wind pressure distribution map, including: Based on the tower structure parameters, the tower is divided into multiple discrete segments along the height direction at preset intervals, and the center height coordinates and windward area of ​​each discrete segment are recorded. Read the wind speed data output by the wind speed sensor at the corresponding height position of each discrete segment as the local wind speed value, and read the wind direction data output by the wind direction sensor; Calculate the difference in local wind speed between adjacent discrete segments. When the difference exceeds a preset gradient threshold, insert a transition segment between the two discrete segments, perform linear interpolation on the wind speed value of the transition segment, and calculate the wind pressure of the transition segment. The bending moment contribution component is obtained by projecting the wind pressure of each discrete segment onto the central axis of the tower. The bending moment contribution components of all discrete segments are accumulated along the height direction to obtain the cumulative bending moment distribution curve. The position with the maximum gradient in the cumulative bending moment distribution curve is identified as the stress concentration height. The center height coordinates, wind pressure, and bending moment contribution components of each discrete segment are constructed into a three-dimensional data array. The discrete segments corresponding to the stress concentration height are marked and arranged in ascending order of center height coordinates to form a wind pressure distribution map containing stress concentration location information.

3. The method according to claim 1, characterized in that, Obtain strain calibration data under no-load and full-load conditions, represent the strain data as a weighted combination of no-load and full-load strain calibration data, solve for the weighting coefficients by minimizing the fitting error, and extract the wind load strain components from the strain data based on the weighting coefficients, including: The crane is controlled to be in no-load and full-load states in a windless environment. The output of the strain sensor is collected and marked as no-load strain calibration data and full-load strain calibration data. Strain data of different combinations of lifting weight and wind speed under actual working conditions were collected. A three-dimensional response surface of strain data, lifting weight and wind speed was established. The strain value corresponding to zero lifting weight on the three-dimensional response surface was extracted as the pure wind load strain benchmark, and the strain value corresponding to zero wind speed was extracted as the pure lifting weight strain benchmark. The correlation coefficient between the pure suspended load strain benchmark and the no-load strain calibration data and the full-load strain calibration data is calculated as the initial weighting coefficient; A search space is constructed with the initial weight coefficients as the center. The weight coefficients in the search space are traversed. The unloaded strain calibration data and the full-load strain calibration data are weighted and combined according to each weight coefficient to obtain the reconstructed strain. The fitting error between the reconstructed strain and the pure suspended strain benchmark is calculated. The weight coefficient with the smallest fitting error is selected as the optimal weight coefficient. The wind load strain components are obtained by weighting the unloaded strain calibration data and the full-load strain calibration data according to the optimal weighting coefficient, and then subtracting the weighted combination result from the strain data.

4. The method according to claim 3, characterized in that, Strain data were collected under actual operating conditions with different combinations of lifting weight and wind speed. A three-dimensional response surface was established based on the strain data and the relationship between lifting weight and wind speed, including: Set the range of load variation and wind speed variation. Within the range of load variation, select multiple load sampling points evenly, and within the range of wind speed variation, select multiple wind speed sampling points evenly. Combine the load sampling points and wind speed sampling points by Cartesian product to obtain the set of sampling working condition points. The crane is controlled to execute the lifting weight and wind speed conditions corresponding to each sampling point in the sampling point set one by one, and the output values ​​of the strain sensor corresponding to each sampling point are collected to establish a mapping relationship between the sampling points and the output values ​​of the strain sensor. The sampling points in the sampling point set that have missing strain sensor output values ​​are marked as missing points. The neighboring points around the missing points that have collected strain sensor output values ​​are extracted. The interpolated strain value of the missing point is calculated by weighted interpolation based on the strain sensor output values ​​of the neighboring points and the distance between the neighboring points and the missing point. The load value of the sampling point is used as the horizontal axis coordinate, the wind speed value of the sampling point is used as the vertical axis coordinate, and the output value of the strain sensor and the interpolated strain value corresponding to the sampling point are used as the vertical axis coordinate to construct a three-dimensional coordinate point cloud. A surface fitting operation is performed on a 3D coordinate point cloud to generate a 3D response surface.

5. The method according to claim 1, characterized in that, The wind pressure distribution map and wind load strain components are input into a pre-trained deep neural network. Based on historical instability accident samples, the similarity between the current state and the instability mode is output, including: The wind pressure distribution map and the wind load strain components are time-aligned to construct a nonlinear coupled response function containing time lag terms, thus obtaining the wind pressure-strain coupled response field. The coupled evolution trajectory sequence is obtained by sliding decomposition of the wind pressure-strain coupled response field along the time dimension. The coupled evolution trajectory sequence is input into the spatial coding subnetwork and temporal coding subnetwork of the deep neural network to extract spatial gradient features and evolution trend features, respectively. The spatial gradient features and evolution trend features are then input into the physical constraint module. The physical constraint module applies consistency constraints based on the mechanical equilibrium relationship between wind pressure and strain to obtain constraint feature vectors. A coupled consistency loss function containing wind pressure-driven error terms and strain inversion error terms is constructed to train a deep neural network, so that the constraint feature vector satisfies the minimum bidirectional prediction error of wind pressure and strain. The current coupled evolution trajectory sequence is dynamically time-warped and matched with the coupled evolution trajectory sequence of historical instability accident samples to calculate the trajectory matching distance. The instability mode with the smallest trajectory matching distance is selected as the target instability mode. Calculate the path deviation and evolution trend consistency coefficient between the current coupled evolution trajectory sequence and the target instability mode trajectory sequence, and calculate the similarity based on the path deviation and evolution trend consistency coefficient.

6. The method according to claim 1, characterized in that, The evolution curve is obtained by nonlinearly fitting the numerical changes of similarity over time. Based on the evolution curve, the critical moments when the similarity reaches the instability threshold are determined, including: The similarity scores are collected continuously at multiple times according to a preset sampling period and arranged in chronological order to form a similarity time series; The similarity time series is segmented by a sliding window and the rate of change of similarity within each segment is calculated. The time and similarity corresponding to the segments with positive rate of change are extracted to form the unstable trend data. The similarity between adjacent sample points in the unstable trend data is subjected to second-order difference operation to obtain a second-order difference value sequence. The absolute value of each second-order difference value in the second-order difference value sequence is normalized and used as the fitting weight coefficient of the corresponding sample point. A polynomial function is used to perform weighted curve fitting on the sample points in the unstable trend data to obtain the initial evolution curve and calculate the fitting residual of each sample point. When the fitting residual exceeds the preset residual threshold, the degree of the polynomial function is increased by a preset step size value and the sample points are re-fitted with weighted curve to obtain the optimized evolution curve. The instability threshold is substituted into the time variable of the optimization evolution curve solution as the critical time when the similarity reaches the instability threshold.

7. The method according to claim 1, characterized in that, The safety margin is calculated based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, an early warning signal and action restriction command are generated, including: Calculate the time difference between the critical moment and the current moment; Obtain the current lifting mass, amplitude value, and slewing angular velocity of the crane. Input the lifting mass, amplitude value, and slewing angular velocity into the historical instability sample library for matching. Extract the actual braking duration of the target historical sample record with the highest matching degree. Subtract the actual braking duration from the time difference to obtain the controllable time window. Obtain the current wind pressure distribution map, input the current wind pressure distribution map into the historical instability sample library for map shape comparison, and count the proportion of samples with wind load abrupt changes in the candidate historical samples with matching shapes as the abrupt change probability. Based on the abrupt change probability, compress the controllable time window to obtain the safety margin. When the safety margin is less than the preset safety value, the safety margin is entered into the historical instability sample library for matching, and the slewing speed limit, lifting speed limit and amplitude speed limit recorded in the reference historical sample with the highest matching degree are extracted. The safety margin and mutation probability are encapsulated to generate an early warning signal. The slewing speed limit, hoisting speed limit, and amplitude speed limit are encapsulated to generate a motion restriction command. The early warning signal is sent to the monitoring terminal, and the motion restriction command is sent to the crane controller.

8. A crane AI monitoring system considering wind load, for implementing the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to collect wind speed and wind direction data through wind speed sensors and wind direction sensors arranged at different height positions on the crane tower, and to collect strain data through strain sensors arranged on the boom. The wind pressure calculation unit is used to divide the tower into multiple discrete segments according to its height, calculate the local wind speed value based on the wind speed data of each discrete segment, and calculate the wind pressure of each discrete segment by combining the windward area and wind resistance coefficient, and arrange them according to spatial location to form a wind pressure distribution map. The strain analysis unit is used to acquire strain calibration data under no-load and full-load conditions. It represents the strain data as a weighted combination of no-load strain calibration data and full-load strain calibration data. The weight coefficients are solved by minimizing the fitting error, and the wind load strain components are extracted from the strain data based on the weight coefficients. The instability prediction unit is used to input the wind pressure distribution map and wind load strain components into a pre-trained deep neural network, and output the similarity between the current state and the instability mode based on historical instability accident samples. The critical judgment unit is used to perform nonlinear fitting on the numerical change of similarity over time to obtain the evolution curve, and to determine the critical moment when the similarity reaches the instability threshold based on the evolution curve. The early warning generation unit is used to calculate the safety margin based on the time difference between the critical moment and the current moment. When the safety margin is less than the preset safety value, it generates an early warning signal and an action restriction command.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images