A crane working monitoring method based on digital twinning

Abstract

The present application relates to the field of data processing, and relates to a crane working monitoring method based on digital twinning, comprising: acquiring real-time operation data of a crane; inputting the real-time operation data of the crane into a physical information neural network embedded with a multi-body dynamics equation of the crane, so as to acquire a high-fidelity state vector representing a real-time working state of the crane; inputting the high-fidelity state vector into a time series graph convolution network, so as to acquire a state sequence trajectory within a future time window; and combining a preset failure mode knowledge base to calculate and generate a multi-dimensional risk vector; if a Mahalanobis distance between a predicted state sequence trajectory and an actual trajectory constituted by subsequent real-time data exceeds a preset threshold, reconstructing parameters of the multi-body dynamics equation embedded in the physical information neural network; the method of the present application can significantly improve the accuracy of crane state monitoring and realize prospective risk prediction.

Description

Technical Field

[0001] The present invention relates to the field of data processing. More specifically, the present invention relates to a crane operation monitoring method based on digital twins. Background Art

[0002] Cranes, as large-scale construction machinery, are used in key sectors such as ports, construction, and manufacturing, where their operational safety and efficiency are paramount. However, cranes are complex structures, face diverse operating conditions, and are often exposed to dynamic loads and harsh environments, making them a high-risk area for safety accidents. Traditional crane safety monitoring relies primarily on operator experience and various independent sensors installed on the equipment, such as torque limiters, travel limiters, anemometers, and collision avoidance devices. These traditional approaches have numerous limitations. First, each sensor system typically operates independently, creating "data silos" that prevent comprehensive assessment of the crane's overall operating status and provide situational awareness. Second, these devices are often threshold-triggered, providing only post-event alarms or passive interventions. They lack the ability to predict potential risks and cannot provide early warning of dynamically evolving risks such as motion interference, sling slack, and structural fatigue. Third, existing safety models are mostly static or simplified empirical models, which struggle to accurately characterize the crane's complex multi-body dynamics under the influence of real-time wind loads, load sway, and mechanism coupling, resulting in insufficient monitoring accuracy and reliability.

[0003] To overcome these limitations, digital twin technology has been introduced into the crane monitoring field, aiming to build high-fidelity virtual models that synchronize with the physical entity and interact with the real world. Currently, there are two main technical approaches to building digital twin models: mechanism-driven and data-driven. Mechanism-driven approaches, such as multibody dynamics simulation or finite element analysis, can accurately describe the physical behavior of cranes, but they are computationally intensive and cannot meet the real-time requirements of monitoring and control. Purely data-driven approaches, such as traditional neural networks, are computationally efficient, but their performance relies heavily on massive amounts of high-quality training data. Data on faults or hazardous conditions, in particular, is extremely difficult to obtain in the real world. Furthermore, the models lack physical constraints, and their predictions may violate fundamental physical laws, raising questions about their generalization and reliability. Furthermore, as physical cranes degrade over time due to wear, fatigue, and material aging, their key physical parameters (such as friction coefficient and component stiffness) change, leading to a gradual mismatch between the pre-established digital twin model and the physical entity.

[0004] In summary, the crane operation monitoring method in the existing technology has technical problems such as insufficient monitoring accuracy and reliability, lack of ability to predict potential risks, and difficulty in meeting the real-time requirements of monitoring and control. Summary of the Invention

[0005] In order to solve the technical problems in the above-mentioned existing crane operation monitoring methods, such as insufficient monitoring accuracy and reliability, lack of ability to predict potential risks, and difficulty in meeting the real-time requirements of monitoring and control, the present invention provides solutions in the following aspects.

[0006] The present invention provides a crane operation monitoring method based on digital twin, comprising:

[0007] Acquire the crane's real-time operating data, including motor current, hook's 3D coordinates, boom pitch angle, load-bearing component strain, and wind speed and direction at the crane's location. This data is then fed into a physical information neural network embedded with the crane's multi-body dynamics equations to generate a high-fidelity state vector representing the crane's real-time operating status.

[0008] The high-fidelity state vector is input into a time-series graph convolutional network to obtain the state sequence trajectory within the future time window. Combined with a preset failure mode knowledge base, a multidimensional risk vector is calculated and generated. The elements of the multidimensional risk vector include: motion interference collision probability, sling relaxation probability, structural fatigue cumulative damage increment, and overturning moment margin.

[0009] If the Mahalanobis distance between the predicted state sequence trajectory and the actual trajectory composed of subsequent real-time data exceeds a preset threshold, a multivariate time series analysis is performed based on the structural causal model to reversely identify the key physical degradation parameters that cause the deviation. The identification results are used to reconstruct the parameters of the multibody dynamics equations embedded in the physical information neural network. The types of key physical degradation parameters include equivalent friction coefficient and component stiffness. The structural causal model is used to describe the causal graph between physical parameters and the prediction errors of each state observation variable. The physical parameters include friction coefficient and component stiffness.

[0010] Preferably, the motor current is obtained through current transformers deployed on the crane's slewing, luffing, lifting and traveling mechanisms; the three-dimensional coordinates of the hook are obtained using the UWB ultra-wideband positioning base station and the hook tag; the pitch angle of the boom is collected using a tilt sensor; strain gauges are attached to the surface of key load-bearing components to measure the strain of the load-bearing components; and a three-dimensional ultrasonic anemometer is installed at the top of the boom to obtain wind speed and direction data.

[0011] Preferably, the training method of the physical information neural network includes:

[0012] A data loss function is constructed based on the difference between the real-time operation data and the high-fidelity state vector output by the neural network, a physical loss function is constructed based on the residual of the crane multi-body dynamics equation, and a weighted summation of the data loss function and the physical loss function is performed to obtain a total loss function;

[0013] The physical information neural network is trained by minimizing the total loss function.

[0014] Preferably, the acquisition of the state sequence trajectory within the future time window includes: defining the crane boom, trolley, hook, load and environmental obstacles within a predetermined range as graph nodes, using the high-fidelity state vector as the node feature, extracting spatial features through a graph convolutional network module, and inputting them into a gated recurrent unit module to learn time dependencies.

[0015] Preferably, the calculation to generate a multidimensional risk vector includes: evaluating the probability of motion interference collision based on the predicted state sequence trajectory; evaluating the probability of cable slack based on the predicted vertical acceleration of the hook; calculating the structural fatigue cumulative damage increment based on the predicted value of component strain and Miner linear cumulative damage criterion; and calculating the overturning moment margin based on the real-time wind speed and load weight.

[0016] Preferably, evaluating the motion interference collision probability includes: evaluating the motion interference collision probability by calculating the minimum predicted distance between the hook or load and the obstacle node based on the predicted state sequence trajectory; the motion interference collision probability is inversely proportional to the minimum predicted distance.

[0017] Preferably, evaluating the slack rope probability based on the predicted vertical acceleration of the hook comprises: calculating the slack rope probability according to the predicted vertical acceleration of the hook, wherein the slack rope probability is proportional to the predicted vertical acceleration of the hook.

[0018] Preferably, calculating the structural fatigue cumulative damage increment includes: analyzing the predicted stress sequence of key component nodes using the rain flow counting method, and calculating the fatigue damage increment based on the SN curve and the Miner cumulative damage criterion.

[0019] Preferably, the step of judging whether the preset threshold is exceeded includes: calculating the Mahalanobis distance between the state vector of the predicted state sequence trajectory and the actual trajectory at the corresponding timestamp, and when the distance is greater than the mean of the Mahalanobis distance calculated based on historical normal operation data plus the standard deviation of a preset multiple, it is judged that the threshold is exceeded.

[0020] Preferably, the reverse identification step includes:

[0021] Determine the key physical degradation parameters that cause the deviation based on a pre-established structural causal model; the structural causal model describes the causal graph between various physical parameters and the prediction errors of various state observation variables; the types of physical parameters include friction coefficient and component stiffness;

[0022] The Bayesian inference algorithm is used to analyze the actual trajectory data that exceeds the threshold and invert the values ​​of the key physical degradation parameters that cause the deviation.

[0023] The beneficial effects of the present invention are as follows: the present invention uses a physical information neural network with embedded multi-body dynamic equations to solve the crane state, deeply integrating the accuracy of the mechanism model and the efficiency of data-driven, and can generate a high-fidelity overall crane state representation that far exceeds the traditional simplified model based on multi-source real-time data, significantly improving the accuracy of state monitoring. In addition, the temporal and spatial evolution of the high-fidelity state is deduced using a time-series graph convolutional network, realizing the transition from passive alarm to forward-looking risk prediction, and can quantitatively evaluate the probability of various complex coupling risks such as motion interference and structural damage occurring in the future time window, providing a richer basis for safety decision-making. Furthermore, a correction closed loop is established between the monitoring model and the physical entity. When the model prediction deviates from the actual operation, it can be traced back and identified to the changes in physical parameters caused by equipment wear, aging, etc., and then the model parameters can be reconstructed to ensure the long-term effectiveness and reliability of the digital twin model throughout the life cycle of the crane. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a flowchart schematically illustrating a crane operation monitoring method based on digital twin according to an embodiment of the present invention;

[0025] Figure 2 is a schematic diagram schematically illustrating a method for collecting real-time operation data of a crane according to an embodiment of the present invention;

[0026] Figure 3 Schematically illustrates the working principle of a temporal graph convolutional network according to an embodiment of the present invention;

[0027] Figure 4 is a schematic diagram schematically illustrating a multi-body dynamics equation parameter reconstruction process according to an embodiment of the present invention;

[0028] Figure 5 Schematic diagram of a physical information neural network and its loss function according to an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Example of crane operation monitoring method based on digital twin:

[0031] like Figure 1 As shown, the crane operation monitoring method based on digital twin of the present invention includes:

[0032] S1. Acquire real-time operating data of the crane, specifically: acquire real-time operating data of the crane, the real-time operating data including motor current, three-dimensional coordinates of the hook, pitch angle of the boom, strain of the load-bearing member, and wind speed and direction at the location of the crane;

[0033] In this embodiment, if Figure 2 As shown, current transformers deployed on the crane's slewing, luffing, hoisting, and traveling mechanisms acquire motor currents. A UWB ultra-wideband positioning base station and a hook tag or GPS module are used to obtain the hook's real-time 3D coordinates. An inclination sensor measures the boom's pitch angle. Strain gauges are attached to the surfaces of key load-bearing components, such as the main beam and boom, to measure strain. A 3D ultrasonic anemometer is installed at the top of the boom to obtain wind speed and direction data. All sensor data is aggregated to an edge computing gateway via a CAN bus or industrial Ethernet for timestamp alignment and normalization preprocessing.

[0034] S2. Obtaining a high-fidelity state vector, specifically: inputting the real-time operating data of the crane into a physical information neural network embedded with the crane's multi-body dynamics equations, thereby obtaining a high-fidelity state vector representing the real-time working state of the crane;

[0035] Specifically, the physical information neural network is a deep feedforward network whose input layer receives preprocessed multidimensional sensor data and whose output layer generates a high-dimensional state vector containing the generalized coordinates, generalized velocities, and generalized accelerations of each crane component. The loss function for network training consists of two weighted components: data loss, which is the mean squared error between specific state quantities output by the network, such as the hook coordinates, and the actual sensor measurements; and physical loss, which involves substituting the complete state vector output by the network into the crane's multibody dynamics equations based on the Lagrange equations and calculating the L2 norm of the residual. The crane's multibody dynamics equations accurately describe the balance between inertia, Coriolis forces, centripetal forces, gravity, and driving forces among the various components, thereby forcing the network output to conform to physical laws.

[0036] S3. Obtaining a state sequence trajectory and a multidimensional risk vector within a future time window, specifically: inputting the high-fidelity state vector into a time-series graph convolutional network to obtain the state sequence trajectory within the future time window; and calculating and generating a multidimensional risk vector based on a preset failure mode knowledge base; the elements of the multidimensional risk vector include: motion interference collision probability, sling relaxation probability, structural fatigue cumulative damage increment, and overturning moment margin;

[0037] Specifically, if Figure 3As shown in the figure, a spatiotemporal graph is constructed, where nodes represent the crane trolley, hook, load, and fixed obstacles within the working range, such as nearby buildings. Node features are the position and velocity information in the high-fidelity state vector, and edges represent the physical connections between components or the spatial proximity between components and obstacles. The sequence of high-fidelity state vectors for consecutive time slices is used as input to a temporal graph convolutional network. The network captures spatial dependencies through graph convolution and temporal dependencies through gated recurrent units or Transformers, outputting predicted state trajectories for each node for dozens of time steps into the future.

[0038] Obtaining the state sequence trajectory within the future time window includes: defining the crane boom, trolley, hook, load and environmental obstacles within a predetermined range as graph nodes, using the high-fidelity state vector as the node feature, extracting spatial features through a graph convolutional network module, and inputting them into a gated recurrent unit module to learn time dependencies.

[0039] Specifically, to effectively predict the future motion trajectories of crane components and loads in complex environments, this method abstracts the entire operating scenario into a dynamic graph structure. In this graph, the crane's boom, trolley, hook, load, and a building column within the operating radius are all defined as independent graph nodes. Each node is accompanied by a feature vector consisting of a high-fidelity state vector. For example, the feature of a hook node may be a vector containing 10 dimensions such as three-dimensional coordinates, velocity, and acceleration. The connections between nodes, that is, the edges of the graph, represent the physical connection or spatial proximity between them.

[0040] When the time series graph convolutional network processes the graph data under this series of time snapshots, the graph convolutional network (GCN) module comes into play first. It updates the feature representation of each node by aggregating information from neighboring nodes, thereby capturing instantaneous spatial relationships. For example, at time t, the GCN can integrate the information of the hook and building column nodes, calculate the spatial proximity between the two, and encode it into new node features. Subsequently, these feature sequences rich in spatial information are fed into the gated recurrent unit (GRU) module. GRU excels at processing time series data. It can learn dependencies in the time dimension, such as the periodicity of the load swing and the inertia of the car's acceleration, and ultimately output a state sequence trajectory prediction for all nodes in the next few seconds.

[0041] Generating a multidimensional risk vector is a quantitative assessment of potential future hazards. In this embodiment, the steps of calculating and generating the multidimensional risk vector include: assessing the probability of motion interference collision based on the predicted state sequence trajectory; assessing the probability of cable slack based on the predicted vertical acceleration of the hook; calculating the structural fatigue cumulative damage increment based on the predicted component strain and Miner's linear cumulative damage criterion; and calculating the overturning moment margin based on the real-time wind speed and load weight.

[0042] In this embodiment, evaluating the motion interference collision probability includes: evaluating the motion interference collision probability by calculating the minimum predicted distance between the hook or load and the obstacle node based on the predicted state sequence trajectory; the motion interference collision probability is inversely proportional to the minimum predicted distance.

[0043] For motion interference collision risks, the system analyzes the predicted trajectory of the load and a known obstacle, such as a 3D model of another tower crane. If the closest distance between the load's outline and the obstacle model decreases to 0.8 meters over the next five seconds, below the set safety threshold of 1.5 meters, a higher collision probability value, such as 0.9, is output.

[0044] In this embodiment, evaluating the slack rope probability based on the predicted vertical acceleration of the hook includes: calculating the slack rope probability according to the predicted vertical acceleration of the hook, and the slack rope probability is proportional to the predicted vertical acceleration of the hook.

[0045] For example, if the downward acceleration is predicted to reach 10 meters per second squared, which exceeds the acceleration due to gravity of 9.8 meters per second squared, it means that the sling is about to lose tension, and a high probability slack risk assessment value is given accordingly.

[0046] The above embodiment evaluates the risk of slack rope by focusing on the predicted acceleration of the hook in the vertical direction. In another embodiment, the probability of slack rope can also be evaluated based on the predicted tension of the rope, including: evaluating the probability of slack rope by checking whether the predicted tension of the rope is less than zero; if the predicted tension of the rope is less than zero, outputting a larger probability of slack rope.

[0047] In this embodiment, calculating the structural fatigue cumulative damage increment includes: analyzing the predicted stress sequence of key component nodes using the rain flow counting method, and calculating the fatigue damage increment based on the SN curve and the Miner cumulative damage criterion.

[0048] Furthermore, for structural fatigue damage, the system applies Miner's linear cumulative damage criterion based on predicted strain data for key boom locations and the material's SN curve. For example, based on the predicted number and amplitude of stress cycles, the system calculates that this operation will increase the cumulative damage at the boom root from 0.352 to 0.354. This increment of 0.0002 becomes a dimension of the risk vector.

[0049] In this embodiment, based on the real-time wind speed and the load weight, calculating the overturning moment margin includes: obtaining the crane's own stabilizing moment , calculate the overturning moment based on the real-time wind speed and load weight combined with the mechanical model ;

[0050] According to the crane's own stabilizing moment and overturning moment Calculating the overturning moment margin , the calculation expression is:

[0051] .

[0052] The capsizing risk assessment combines real-time wind speed data and load weight. Assuming a wind speed of 20 meters per second and a load of 30 tons, the system calculates a capsizing moment of 4,500 kNm based on the mechanical model. The inherent stabilizing moment of the equipment is 6,000 kNm, resulting in a capsizing moment margin of 25%. This value serves as the final dimension of the risk vector, comprehensively characterizing the immediate and long-term risks of the current operation.

[0053] S4, reconstructing the multi-body dynamics equation parameters embedded in the physical information neural network, specifically: Figure 4 As shown, if the Mahalanobis distance between the predicted state sequence trajectory and the actual trajectory composed of subsequent real-time data exceeds a preset threshold, a multivariate time series analysis is performed based on the structural causal model to reversely identify the key physical degradation parameters that cause the deviation. The identification results are used to reconstruct the parameters of the multibody dynamics equations embedded in the physical information neural network. The types of key physical degradation parameters include equivalent friction coefficient and component stiffness. The structural causal model is used to describe the causal graph between physical parameters and the prediction errors of each state observation variable. Physical parameters include friction coefficient and component stiffness.

[0054] Specifically, the Mahalanobis distance between the state vector sequence predicted by the temporal graph convolutional network and the actual state vector sequence subsequently generated by the physical information neural network is continuously calculated. This distance can account for the correlation between state variables. Once the distance exceeds a 99 percent confidence upper threshold based on historical data statistics, the identification process is triggered.

[0055] In this embodiment, the step of determining whether the preset threshold is exceeded includes: calculating the Mahalanobis distance between the state vector of the predicted state sequence trajectory and the actual trajectory at the corresponding timestamp, and when the distance is greater than the mean of the Mahalanobis distance calculated based on historical normal operation data plus the standard deviation of a preset multiple, it is determined that the threshold is exceeded.

[0056] Using the Mahalanobis distance to quantify the deviation between prediction and reality effectively accounts for the correlation between variables of different dimensions in the state vector, making it more accurate than the Euclidean distance. For example, the speed of a crane trolley and its motor current are highly correlated, and the Mahalanobis distance takes this inherent connection into account when calculating the deviation. At each point in time, the system obtains a predicted state vector (e.g., a predicted trolley speed of 1.5 meters per second and a motor current of 30 amps) and the actual sensor measurements (e.g., a speed of 1.2 meters per second and a current of 25 amps). The Mahalanobis distance between these two vectors is then calculated using the covariance matrix.

[0057] The judgment threshold is dynamically generated through statistical learning. Thousands of hours of historical normal operation data are analyzed to calculate the Mahalanobis distance distribution between the predicted and actual trajectories under these normal operating conditions. Assume that through statistical analysis, the average Mahalanobis distance of normal deviation is 2.0 and the standard deviation is 0.4. If the preset multiplier is set to 3, that is, the three sigma criterion is adopted, the judgment threshold is set to 2.0 plus 3 times 0.4, which is 3.2. When a real-time calculated Mahalanobis distance, such as 3.5, exceeds this threshold of 3.2, the system determines that a significant abnormal deviation has occurred, indicating that there is a non-negligible mismatch between the digital twin and the physical entity.

[0058] In one embodiment, the reverse identification step includes:

[0059] S301. Determine key physical degradation parameters that cause deviations based on a pre-established structural causal model; the structural causal model describes a causal graph between various physical parameters and prediction errors of various state observation variables; the types of physical parameters include friction coefficient and component stiffness;

[0060] Specifically, when a deviation between the predicted and actual trajectory exceeds a threshold, a reverse identification process is initiated to explore the root cause. At the heart of this process is a pre-built structural causal model, which depicts the causal chain in the physical world using a directed acyclic graph. Based on physical principles and extensive experimental data, the model distinguishes between correlation and causality, providing a foundation for accurate tracing.

[0061] For example, the model defines that an increase in the equivalent friction coefficient of the track will result in a decrease in the actual vehicle acceleration and an increase in the motor drive current at the same drive voltage. Once actual operating data exceeding a threshold is obtained, such as observing that the vehicle acceleration is 0.2 meters per second squared lower than expected, the Bayesian inference algorithm begins to work.

[0062] In this embodiment, a structural causal model is constructed using the PC algorithm or the LiNGAM algorithm. This model describes a causal graph between physical parameters such as friction coefficient and component stiffness and the prediction errors of various state observation variables. By analyzing the Granger causality and partial correlations between the prediction error time series, the root physical parameters that contribute most to the current model deviation are identified.

[0063] Using the anomalous observational data as evidence, the probability distribution of potential degradation parameters in the structural causal model, such as the friction coefficient and component stiffness, is updated. Given the current observational data, the probability of a 20% increase in the friction coefficient or a 10% decrease in the stiffness at the base of the boom is calculated. If the calculation results show a posterior probability of a friction coefficient increase of 0.85, while the posterior probabilities of changes in other parameters are all below 0.1, the abnormal increase in the equivalent friction coefficient is determined to be the key physical degradation parameter causing the deviation.

[0064] S302: Analyze the actual trajectory data exceeding the threshold using a Bayesian inference algorithm to invert the values ​​of key physical degradation parameters that cause the deviation.

[0065] For example, if the prediction error for the trolley's motion shows a strong correlation with the prediction error for the hoisting motion, and causal analysis points to the coupling stiffness, a Bayesian optimization algorithm is activated to search for the optimal equivalent coupling stiffness, with the goal of minimizing the error between the model output and the actual data. This value is then used as the identified stiffness value. This identified stiffness value is then updated into the multibody dynamics equations embedded in the physical information neural network, completing the online correction of the model.

[0066] In the previous step, reverse identification revealed changes in key physical parameters that caused the model deviation. For example, the equivalent friction coefficient of the trolley track was found to have increased from the initial value of 0.05 due to wear to 0.08. This newly identified value reflects the current health status of the physical equipment.

[0067] The new friction coefficient value of 0.08 is updated to the multi-body dynamics equations that the physical information neural network relies on. Originally, when calculating the physical loss function, the equation for the motion of the car used the old friction coefficient value, but now it is replaced by this more accurate current value. After the parameter update is completed, the neural network can be fine-tuned and trained. Because the embedded physical model is closer to reality, the prediction accuracy of the model will be significantly improved. For example, the prediction error of the car's position may be reduced from an average of 5 cm to 1 cm. In this way, the digital twin can continuously and automatically calibrate itself, so that it can maintain high fidelity throughout the life cycle of the device.

[0068] like Figure 5 As shown, in one embodiment, the training method of the physical information neural network includes:

[0069] S101, constructing a data loss function based on the difference between the real-time operation data and the high-fidelity state vector output by the neural network, constructing a physical loss function based on the residual of the crane multi-body dynamics equation, and performing a weighted summation of the data loss function and the physical loss function to obtain a total loss function;

[0070] S102: Training the physical information neural network by minimizing the total loss function. This includes:

[0071] a) Input real-time operation data into the physical information neural network to obtain a high-fidelity state vector;

[0072] b. Calculate the total loss function based on the high-fidelity state vector;

[0073] c. Adjust the parameters within the network;

[0074] The parameters within the network may include weights and biases, and in other embodiments, may also include other parameters.

[0075] d Iteratively calculate the total loss function until the total loss function reaches the minimum value and stop training.

[0076] Continuously adjust the parameters within the network to make the total loss function as small as possible. When the total loss function reaches the minimum, stop training.

[0077] To ensure that the generated digital twin state vector not only matches the real-world measurement data but also adheres to the basic laws of rigid-body motion, a training strategy driven by both data and physics is adopted. The data loss function is constructed based on the difference between the real-time data collected by the sensor and the value predicted by the neural network. For example, if the sensor measurement of the car's position at a certain moment is 15.2 meters, and the neural network output is 15.1 meters, there is a difference of 0.1 meters. The data loss term is formed by accumulating the mean squared error of thousands of such data points across all state dimensions. Its goal is to make the model output as close as possible to real-world observations.

[0078] At the same time, the construction of the physical loss function is to substitute the state vector output by the neural network, such as displacement, velocity and acceleration, into the known differential equation of crane multi-body dynamics. For example, for the swing of the hook, its dynamic equation can be expressed as F equals ma. If the acceleration a output by the neural network and the calculated force F are substituted, the two sides of the equation are not equal, and the difference constitutes the physical residual. The physical loss term is the sum of the squares of these residuals, and its goal is to force the learning results of the neural network not to violate physical constraints such as Newton's second law. Finally, by setting the weight coefficients, such as a data loss weight of 0.6 and a physical loss weight of 0.4, the weighted sum of the two is used to obtain the total loss function, and the gradient descent algorithm is used for minimization training to obtain a high-fidelity physical information neural network model.

[0079] In one embodiment, a method for obtaining a crane multi-body dynamics equation includes:

[0080] S201. Construct a crane multi-body system model with multiple rigid bodies;

[0081] The rigid bodies of the crane multi-body system model include the base, boom, pulley, hook, etc. The motion types of the crane multi-body system model include rotation, translation, extension, etc.

[0082] S202, establishing generalized coordinates;

[0083] Select the independent degrees of freedom of the crane multibody system. In this embodiment, the degrees of freedom are 5.

[0084] The generalized coordinates established include:

[0085] q1: base rotation angle;

[0086] q2: boom elevation angle;

[0087] q3: telescopic length of the boom;

[0088] q4: Hook drop height;

[0089] q5: Displacement of the trolley on the slide rail (such as a bridge crane).

[0090] S203, obtaining the total kinetic energy expression of the crane multi-body system;

[0091] Add up the translational kinetic energy and rotational kinetic energy of each rigid body and the translational kinetic energy and rotational kinetic energy of all rigid bodies to obtain the total kinetic energy of the crane multi-body system.

[0092] S204. Obtain the total potential energy expression of the crane multi-body system;

[0093] S205. Calculate the generalized forces of the crane multi-body system;

[0094] S206. For each generalized coordinate, substitute the expressions for total kinetic energy and total potential energy, and then substitute them into the Lagrange equation to solve the dynamic equation in the direction of that coordinate.

[0095] For the i-th generalized coordinate , the Lagrange equation is:

[0096]

[0097] Where T is the total kinetic energy of the crane multi-body system, V is the total potential energy of the crane multi-body system, and q i is the ith generalized coordinate, is the ith generalized velocity, is a non-conservative generalized force, and t represents time.

Claims

1. A crane operation monitoring method based on digital twin, characterized in that: include: Acquire the crane's real-time operating data, including motor current, hook's 3D coordinates, boom pitch angle, load-bearing component strain, and wind speed and direction at the crane's location. This data is then fed into a physical information neural network embedded with the crane's multi-body dynamics equations to generate a high-fidelity state vector representing the crane's real-time operating status. Inputting the high-fidelity state vector into a temporal graph convolutional network to obtain a state sequence trajectory in a future time window; Combined with the preset failure mode knowledge base, a multi-dimensional risk vector is calculated and generated; The elements of the multi-dimensional risk vector include: motion interference collision probability, sling relaxation probability, structural fatigue cumulative damage increment and overturning moment margin; If the Mahalanobis distance between the predicted state sequence trajectory and the actual trajectory composed of subsequent real-time data exceeds a preset threshold, a multivariate time series analysis is performed based on the structural causal model to reversely identify the key physical degradation parameters that cause the deviation. The identification results are used to reconstruct the parameters of the multibody dynamics equations embedded in the physical information neural network. The types of key physical degradation parameters include equivalent friction coefficient and component stiffness. The structural causal model is used to describe the causal graph between physical parameters and the prediction errors of each state observation variable. The physical parameters include friction coefficient and component stiffness.

2. The crane operation monitoring method based on digital twin according to claim 1, characterized in that: The motor current is obtained through current transformers deployed on the crane's slewing, luffing, lifting and traveling mechanisms; the three-dimensional coordinates of the hook are obtained using the UWB ultra-wideband positioning base station and the hook tag; the pitch angle of the boom is collected using a tilt sensor; strain gauges are attached to the surface of key load-bearing components to measure the strain of the load-bearing components; and a three-dimensional ultrasonic anemometer is installed at the top of the boom to obtain wind speed and direction data.

3. The crane operation monitoring method based on digital twin according to claim 1, characterized in that: The training method of the physical information neural network includes: A data loss function is constructed based on the difference between the real-time operation data and the high-fidelity state vector output by the neural network, a physical loss function is constructed based on the residual of the crane multi-body dynamics equation, and a weighted summation of the data loss function and the physical loss function is performed to obtain a total loss function; The physical information neural network is trained by minimizing the total loss function.

4. The crane operation monitoring method based on digital twin according to claim 1, characterized in that: The method of obtaining the state sequence trajectory within the future time window includes: defining the crane boom, trolley, hook, load and environmental obstacles within a predetermined range as graph nodes, using the high-fidelity state vector as the node feature, extracting spatial features through a graph convolutional network module, and inputting them into a gated recurrent unit module to learn time dependencies.

5. The crane operation monitoring method based on digital twin according to claim 1, characterized in that: The calculation and generation of the multidimensional risk vector includes: evaluating the probability of motion interference collision based on the predicted state sequence trajectory; evaluating the probability of cable relaxation based on the predicted vertical acceleration of the hook; calculating the structural fatigue cumulative damage increment based on the predicted value of component strain and Miner linear cumulative damage criterion; and calculating the overturning moment margin based on the real-time wind speed and load weight.

6. The crane operation monitoring method based on digital twin according to claim 5, characterized in that: Evaluating the motion interference collision probability includes: evaluating the motion interference collision probability by calculating the minimum predicted distance between the hook or load and the obstacle node based on the predicted state sequence trajectory; the motion interference collision probability is inversely proportional to the minimum predicted distance.

7. The crane operation monitoring method based on digital twin according to claim 5, characterized in that: Evaluating the slack probability of the sling based on the predicted vertical acceleration of the hook includes: calculating the slack probability of the sling according to the predicted vertical acceleration of the hook, and the slack probability of the sling is proportional to the predicted vertical acceleration of the hook.

8. The crane operation monitoring method based on digital twin according to claim 5, characterized in that: Calculating the structural fatigue cumulative damage increment includes: analyzing the predicted stress sequence of key component nodes using the rain flow counting method, and calculating the fatigue damage increment based on the SN curve and Miner cumulative damage criterion.

9. The crane operation monitoring method based on digital twin according to claim 1, characterized in that: The step of judging whether the preset threshold is exceeded includes: calculating the Mahalanobis distance between the state vectors of the predicted state sequence trajectory and the actual trajectory at the corresponding timestamp, and judging that the threshold is exceeded when the distance is greater than the mean of the Mahalanobis distance calculated based on historical normal operation data plus a preset multiple of the standard deviation.

10. The crane operation monitoring method based on digital twin according to any one of claims 1 to 9, characterized in that: The reverse identification step includes: Determine the key physical degradation parameters that cause the deviation based on a pre-established structural causal model; the structural causal model describes the causal graph between various physical parameters and the prediction errors of various state observation variables; the types of physical parameters include friction coefficient and component stiffness; The Bayesian inference algorithm is used to analyze the actual trajectory data that exceeds the threshold and invert the values ​​of the key physical degradation parameters that cause the deviation.

Images